annotate src/share/vm/opto/compile.cpp @ 2733:c124e2e7463e

7083786: dead various dead chunks of code Reviewed-by: iveresov, kvn
author never
date Wed, 31 Aug 2011 16:46:11 -0700
parents 11211f7cb5a0
children 670a74b863fc
rev   line source
duke@0 1 /*
never@2223 2 * Copyright (c) 1997, 2011, Oracle and/or its affiliates. All rights reserved.
duke@0 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
duke@0 4 *
duke@0 5 * This code is free software; you can redistribute it and/or modify it
duke@0 6 * under the terms of the GNU General Public License version 2 only, as
duke@0 7 * published by the Free Software Foundation.
duke@0 8 *
duke@0 9 * This code is distributed in the hope that it will be useful, but WITHOUT
duke@0 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
duke@0 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
duke@0 12 * version 2 for more details (a copy is included in the LICENSE file that
duke@0 13 * accompanied this code).
duke@0 14 *
duke@0 15 * You should have received a copy of the GNU General Public License version
duke@0 16 * 2 along with this work; if not, write to the Free Software Foundation,
duke@0 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
duke@0 18 *
trims@1472 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
trims@1472 20 * or visit www.oracle.com if you need additional information or have any
trims@1472 21 * questions.
duke@0 22 *
duke@0 23 */
duke@0 24
stefank@1879 25 #include "precompiled.hpp"
stefank@1879 26 #include "asm/assembler.hpp"
stefank@1879 27 #include "classfile/systemDictionary.hpp"
stefank@1879 28 #include "code/exceptionHandlerTable.hpp"
stefank@1879 29 #include "code/nmethod.hpp"
stefank@1879 30 #include "compiler/compileLog.hpp"
stefank@1879 31 #include "compiler/oopMap.hpp"
stefank@1879 32 #include "opto/addnode.hpp"
stefank@1879 33 #include "opto/block.hpp"
stefank@1879 34 #include "opto/c2compiler.hpp"
stefank@1879 35 #include "opto/callGenerator.hpp"
stefank@1879 36 #include "opto/callnode.hpp"
stefank@1879 37 #include "opto/cfgnode.hpp"
stefank@1879 38 #include "opto/chaitin.hpp"
stefank@1879 39 #include "opto/compile.hpp"
stefank@1879 40 #include "opto/connode.hpp"
stefank@1879 41 #include "opto/divnode.hpp"
stefank@1879 42 #include "opto/escape.hpp"
stefank@1879 43 #include "opto/idealGraphPrinter.hpp"
stefank@1879 44 #include "opto/loopnode.hpp"
stefank@1879 45 #include "opto/machnode.hpp"
stefank@1879 46 #include "opto/macro.hpp"
stefank@1879 47 #include "opto/matcher.hpp"
stefank@1879 48 #include "opto/memnode.hpp"
stefank@1879 49 #include "opto/mulnode.hpp"
stefank@1879 50 #include "opto/node.hpp"
stefank@1879 51 #include "opto/opcodes.hpp"
stefank@1879 52 #include "opto/output.hpp"
stefank@1879 53 #include "opto/parse.hpp"
stefank@1879 54 #include "opto/phaseX.hpp"
stefank@1879 55 #include "opto/rootnode.hpp"
stefank@1879 56 #include "opto/runtime.hpp"
stefank@1879 57 #include "opto/stringopts.hpp"
stefank@1879 58 #include "opto/type.hpp"
stefank@1879 59 #include "opto/vectornode.hpp"
stefank@1879 60 #include "runtime/arguments.hpp"
stefank@1879 61 #include "runtime/signature.hpp"
stefank@1879 62 #include "runtime/stubRoutines.hpp"
stefank@1879 63 #include "runtime/timer.hpp"
stefank@1879 64 #include "utilities/copy.hpp"
stefank@1879 65 #ifdef TARGET_ARCH_MODEL_x86_32
stefank@1879 66 # include "adfiles/ad_x86_32.hpp"
stefank@1879 67 #endif
stefank@1879 68 #ifdef TARGET_ARCH_MODEL_x86_64
stefank@1879 69 # include "adfiles/ad_x86_64.hpp"
stefank@1879 70 #endif
stefank@1879 71 #ifdef TARGET_ARCH_MODEL_sparc
stefank@1879 72 # include "adfiles/ad_sparc.hpp"
stefank@1879 73 #endif
stefank@1879 74 #ifdef TARGET_ARCH_MODEL_zero
stefank@1879 75 # include "adfiles/ad_zero.hpp"
stefank@1879 76 #endif
bobv@2073 77 #ifdef TARGET_ARCH_MODEL_arm
bobv@2073 78 # include "adfiles/ad_arm.hpp"
bobv@2073 79 #endif
bobv@2073 80 #ifdef TARGET_ARCH_MODEL_ppc
bobv@2073 81 # include "adfiles/ad_ppc.hpp"
bobv@2073 82 #endif
duke@0 83
twisti@1915 84
twisti@1915 85 // -------------------- Compile::mach_constant_base_node -----------------------
twisti@1915 86 // Constant table base node singleton.
twisti@1915 87 MachConstantBaseNode* Compile::mach_constant_base_node() {
twisti@1915 88 if (_mach_constant_base_node == NULL) {
twisti@1915 89 _mach_constant_base_node = new (C) MachConstantBaseNode();
twisti@1915 90 _mach_constant_base_node->add_req(C->root());
twisti@1915 91 }
twisti@1915 92 return _mach_constant_base_node;
twisti@1915 93 }
twisti@1915 94
twisti@1915 95
duke@0 96 /// Support for intrinsics.
duke@0 97
duke@0 98 // Return the index at which m must be inserted (or already exists).
duke@0 99 // The sort order is by the address of the ciMethod, with is_virtual as minor key.
duke@0 100 int Compile::intrinsic_insertion_index(ciMethod* m, bool is_virtual) {
duke@0 101 #ifdef ASSERT
duke@0 102 for (int i = 1; i < _intrinsics->length(); i++) {
duke@0 103 CallGenerator* cg1 = _intrinsics->at(i-1);
duke@0 104 CallGenerator* cg2 = _intrinsics->at(i);
duke@0 105 assert(cg1->method() != cg2->method()
duke@0 106 ? cg1->method() < cg2->method()
duke@0 107 : cg1->is_virtual() < cg2->is_virtual(),
duke@0 108 "compiler intrinsics list must stay sorted");
duke@0 109 }
duke@0 110 #endif
duke@0 111 // Binary search sorted list, in decreasing intervals [lo, hi].
duke@0 112 int lo = 0, hi = _intrinsics->length()-1;
duke@0 113 while (lo <= hi) {
duke@0 114 int mid = (uint)(hi + lo) / 2;
duke@0 115 ciMethod* mid_m = _intrinsics->at(mid)->method();
duke@0 116 if (m < mid_m) {
duke@0 117 hi = mid-1;
duke@0 118 } else if (m > mid_m) {
duke@0 119 lo = mid+1;
duke@0 120 } else {
duke@0 121 // look at minor sort key
duke@0 122 bool mid_virt = _intrinsics->at(mid)->is_virtual();
duke@0 123 if (is_virtual < mid_virt) {
duke@0 124 hi = mid-1;
duke@0 125 } else if (is_virtual > mid_virt) {
duke@0 126 lo = mid+1;
duke@0 127 } else {
duke@0 128 return mid; // exact match
duke@0 129 }
duke@0 130 }
duke@0 131 }
duke@0 132 return lo; // inexact match
duke@0 133 }
duke@0 134
duke@0 135 void Compile::register_intrinsic(CallGenerator* cg) {
duke@0 136 if (_intrinsics == NULL) {
duke@0 137 _intrinsics = new GrowableArray<CallGenerator*>(60);
duke@0 138 }
duke@0 139 // This code is stolen from ciObjectFactory::insert.
duke@0 140 // Really, GrowableArray should have methods for
duke@0 141 // insert_at, remove_at, and binary_search.
duke@0 142 int len = _intrinsics->length();
duke@0 143 int index = intrinsic_insertion_index(cg->method(), cg->is_virtual());
duke@0 144 if (index == len) {
duke@0 145 _intrinsics->append(cg);
duke@0 146 } else {
duke@0 147 #ifdef ASSERT
duke@0 148 CallGenerator* oldcg = _intrinsics->at(index);
duke@0 149 assert(oldcg->method() != cg->method() || oldcg->is_virtual() != cg->is_virtual(), "don't register twice");
duke@0 150 #endif
duke@0 151 _intrinsics->append(_intrinsics->at(len-1));
duke@0 152 int pos;
duke@0 153 for (pos = len-2; pos >= index; pos--) {
duke@0 154 _intrinsics->at_put(pos+1,_intrinsics->at(pos));
duke@0 155 }
duke@0 156 _intrinsics->at_put(index, cg);
duke@0 157 }
duke@0 158 assert(find_intrinsic(cg->method(), cg->is_virtual()) == cg, "registration worked");
duke@0 159 }
duke@0 160
duke@0 161 CallGenerator* Compile::find_intrinsic(ciMethod* m, bool is_virtual) {
duke@0 162 assert(m->is_loaded(), "don't try this on unloaded methods");
duke@0 163 if (_intrinsics != NULL) {
duke@0 164 int index = intrinsic_insertion_index(m, is_virtual);
duke@0 165 if (index < _intrinsics->length()
duke@0 166 && _intrinsics->at(index)->method() == m
duke@0 167 && _intrinsics->at(index)->is_virtual() == is_virtual) {
duke@0 168 return _intrinsics->at(index);
duke@0 169 }
duke@0 170 }
duke@0 171 // Lazily create intrinsics for intrinsic IDs well-known in the runtime.
jrose@856 172 if (m->intrinsic_id() != vmIntrinsics::_none &&
jrose@856 173 m->intrinsic_id() <= vmIntrinsics::LAST_COMPILER_INLINE) {
duke@0 174 CallGenerator* cg = make_vm_intrinsic(m, is_virtual);
duke@0 175 if (cg != NULL) {
duke@0 176 // Save it for next time:
duke@0 177 register_intrinsic(cg);
duke@0 178 return cg;
duke@0 179 } else {
duke@0 180 gather_intrinsic_statistics(m->intrinsic_id(), is_virtual, _intrinsic_disabled);
duke@0 181 }
duke@0 182 }
duke@0 183 return NULL;
duke@0 184 }
duke@0 185
duke@0 186 // Compile:: register_library_intrinsics and make_vm_intrinsic are defined
duke@0 187 // in library_call.cpp.
duke@0 188
duke@0 189
duke@0 190 #ifndef PRODUCT
duke@0 191 // statistics gathering...
duke@0 192
duke@0 193 juint Compile::_intrinsic_hist_count[vmIntrinsics::ID_LIMIT] = {0};
duke@0 194 jubyte Compile::_intrinsic_hist_flags[vmIntrinsics::ID_LIMIT] = {0};
duke@0 195
duke@0 196 bool Compile::gather_intrinsic_statistics(vmIntrinsics::ID id, bool is_virtual, int flags) {
duke@0 197 assert(id > vmIntrinsics::_none && id < vmIntrinsics::ID_LIMIT, "oob");
duke@0 198 int oflags = _intrinsic_hist_flags[id];
duke@0 199 assert(flags != 0, "what happened?");
duke@0 200 if (is_virtual) {
duke@0 201 flags |= _intrinsic_virtual;
duke@0 202 }
duke@0 203 bool changed = (flags != oflags);
duke@0 204 if ((flags & _intrinsic_worked) != 0) {
duke@0 205 juint count = (_intrinsic_hist_count[id] += 1);
duke@0 206 if (count == 1) {
duke@0 207 changed = true; // first time
duke@0 208 }
duke@0 209 // increment the overall count also:
duke@0 210 _intrinsic_hist_count[vmIntrinsics::_none] += 1;
duke@0 211 }
duke@0 212 if (changed) {
duke@0 213 if (((oflags ^ flags) & _intrinsic_virtual) != 0) {
duke@0 214 // Something changed about the intrinsic's virtuality.
duke@0 215 if ((flags & _intrinsic_virtual) != 0) {
duke@0 216 // This is the first use of this intrinsic as a virtual call.
duke@0 217 if (oflags != 0) {
duke@0 218 // We already saw it as a non-virtual, so note both cases.
duke@0 219 flags |= _intrinsic_both;
duke@0 220 }
duke@0 221 } else if ((oflags & _intrinsic_both) == 0) {
duke@0 222 // This is the first use of this intrinsic as a non-virtual
duke@0 223 flags |= _intrinsic_both;
duke@0 224 }
duke@0 225 }
duke@0 226 _intrinsic_hist_flags[id] = (jubyte) (oflags | flags);
duke@0 227 }
duke@0 228 // update the overall flags also:
duke@0 229 _intrinsic_hist_flags[vmIntrinsics::_none] |= (jubyte) flags;
duke@0 230 return changed;
duke@0 231 }
duke@0 232
duke@0 233 static char* format_flags(int flags, char* buf) {
duke@0 234 buf[0] = 0;
duke@0 235 if ((flags & Compile::_intrinsic_worked) != 0) strcat(buf, ",worked");
duke@0 236 if ((flags & Compile::_intrinsic_failed) != 0) strcat(buf, ",failed");
duke@0 237 if ((flags & Compile::_intrinsic_disabled) != 0) strcat(buf, ",disabled");
duke@0 238 if ((flags & Compile::_intrinsic_virtual) != 0) strcat(buf, ",virtual");
duke@0 239 if ((flags & Compile::_intrinsic_both) != 0) strcat(buf, ",nonvirtual");
duke@0 240 if (buf[0] == 0) strcat(buf, ",");
duke@0 241 assert(buf[0] == ',', "must be");
duke@0 242 return &buf[1];
duke@0 243 }
duke@0 244
duke@0 245 void Compile::print_intrinsic_statistics() {
duke@0 246 char flagsbuf[100];
duke@0 247 ttyLocker ttyl;
duke@0 248 if (xtty != NULL) xtty->head("statistics type='intrinsic'");
duke@0 249 tty->print_cr("Compiler intrinsic usage:");
duke@0 250 juint total = _intrinsic_hist_count[vmIntrinsics::_none];
duke@0 251 if (total == 0) total = 1; // avoid div0 in case of no successes
duke@0 252 #define PRINT_STAT_LINE(name, c, f) \
duke@0 253 tty->print_cr(" %4d (%4.1f%%) %s (%s)", (int)(c), ((c) * 100.0) / total, name, f);
duke@0 254 for (int index = 1 + (int)vmIntrinsics::_none; index < (int)vmIntrinsics::ID_LIMIT; index++) {
duke@0 255 vmIntrinsics::ID id = (vmIntrinsics::ID) index;
duke@0 256 int flags = _intrinsic_hist_flags[id];
duke@0 257 juint count = _intrinsic_hist_count[id];
duke@0 258 if ((flags | count) != 0) {
duke@0 259 PRINT_STAT_LINE(vmIntrinsics::name_at(id), count, format_flags(flags, flagsbuf));
duke@0 260 }
duke@0 261 }
duke@0 262 PRINT_STAT_LINE("total", total, format_flags(_intrinsic_hist_flags[vmIntrinsics::_none], flagsbuf));
duke@0 263 if (xtty != NULL) xtty->tail("statistics");
duke@0 264 }
duke@0 265
duke@0 266 void Compile::print_statistics() {
duke@0 267 { ttyLocker ttyl;
duke@0 268 if (xtty != NULL) xtty->head("statistics type='opto'");
duke@0 269 Parse::print_statistics();
duke@0 270 PhaseCCP::print_statistics();
duke@0 271 PhaseRegAlloc::print_statistics();
duke@0 272 Scheduling::print_statistics();
duke@0 273 PhasePeephole::print_statistics();
duke@0 274 PhaseIdealLoop::print_statistics();
duke@0 275 if (xtty != NULL) xtty->tail("statistics");
duke@0 276 }
duke@0 277 if (_intrinsic_hist_flags[vmIntrinsics::_none] != 0) {
duke@0 278 // put this under its own <statistics> element.
duke@0 279 print_intrinsic_statistics();
duke@0 280 }
duke@0 281 }
duke@0 282 #endif //PRODUCT
duke@0 283
duke@0 284 // Support for bundling info
duke@0 285 Bundle* Compile::node_bundling(const Node *n) {
duke@0 286 assert(valid_bundle_info(n), "oob");
duke@0 287 return &_node_bundling_base[n->_idx];
duke@0 288 }
duke@0 289
duke@0 290 bool Compile::valid_bundle_info(const Node *n) {
duke@0 291 return (_node_bundling_limit > n->_idx);
duke@0 292 }
duke@0 293
duke@0 294
never@1080 295 void Compile::gvn_replace_by(Node* n, Node* nn) {
never@1080 296 for (DUIterator_Last imin, i = n->last_outs(imin); i >= imin; ) {
never@1080 297 Node* use = n->last_out(i);
never@1080 298 bool is_in_table = initial_gvn()->hash_delete(use);
never@1080 299 uint uses_found = 0;
never@1080 300 for (uint j = 0; j < use->len(); j++) {
never@1080 301 if (use->in(j) == n) {
never@1080 302 if (j < use->req())
never@1080 303 use->set_req(j, nn);
never@1080 304 else
never@1080 305 use->set_prec(j, nn);
never@1080 306 uses_found++;
never@1080 307 }
never@1080 308 }
never@1080 309 if (is_in_table) {
never@1080 310 // reinsert into table
never@1080 311 initial_gvn()->hash_find_insert(use);
never@1080 312 }
never@1080 313 record_for_igvn(use);
never@1080 314 i -= uses_found; // we deleted 1 or more copies of this edge
never@1080 315 }
never@1080 316 }
never@1080 317
never@1080 318
never@1080 319
never@1080 320
duke@0 321 // Identify all nodes that are reachable from below, useful.
duke@0 322 // Use breadth-first pass that records state in a Unique_Node_List,
duke@0 323 // recursive traversal is slower.
duke@0 324 void Compile::identify_useful_nodes(Unique_Node_List &useful) {
duke@0 325 int estimated_worklist_size = unique();
duke@0 326 useful.map( estimated_worklist_size, NULL ); // preallocate space
duke@0 327
duke@0 328 // Initialize worklist
duke@0 329 if (root() != NULL) { useful.push(root()); }
duke@0 330 // If 'top' is cached, declare it useful to preserve cached node
duke@0 331 if( cached_top_node() ) { useful.push(cached_top_node()); }
duke@0 332
duke@0 333 // Push all useful nodes onto the list, breadthfirst
duke@0 334 for( uint next = 0; next < useful.size(); ++next ) {
duke@0 335 assert( next < unique(), "Unique useful nodes < total nodes");
duke@0 336 Node *n = useful.at(next);
duke@0 337 uint max = n->len();
duke@0 338 for( uint i = 0; i < max; ++i ) {
duke@0 339 Node *m = n->in(i);
duke@0 340 if( m == NULL ) continue;
duke@0 341 useful.push(m);
duke@0 342 }
duke@0 343 }
duke@0 344 }
duke@0 345
duke@0 346 // Disconnect all useless nodes by disconnecting those at the boundary.
duke@0 347 void Compile::remove_useless_nodes(Unique_Node_List &useful) {
duke@0 348 uint next = 0;
duke@0 349 while( next < useful.size() ) {
duke@0 350 Node *n = useful.at(next++);
duke@0 351 // Use raw traversal of out edges since this code removes out edges
duke@0 352 int max = n->outcnt();
duke@0 353 for (int j = 0; j < max; ++j ) {
duke@0 354 Node* child = n->raw_out(j);
duke@0 355 if( ! useful.member(child) ) {
duke@0 356 assert( !child->is_top() || child != top(),
duke@0 357 "If top is cached in Compile object it is in useful list");
duke@0 358 // Only need to remove this out-edge to the useless node
duke@0 359 n->raw_del_out(j);
duke@0 360 --j;
duke@0 361 --max;
duke@0 362 }
duke@0 363 }
duke@0 364 if (n->outcnt() == 1 && n->has_special_unique_user()) {
duke@0 365 record_for_igvn( n->unique_out() );
duke@0 366 }
duke@0 367 }
duke@0 368 debug_only(verify_graph_edges(true/*check for no_dead_code*/);)
duke@0 369 }
duke@0 370
duke@0 371 //------------------------------frame_size_in_words-----------------------------
duke@0 372 // frame_slots in units of words
duke@0 373 int Compile::frame_size_in_words() const {
duke@0 374 // shift is 0 in LP32 and 1 in LP64
duke@0 375 const int shift = (LogBytesPerWord - LogBytesPerInt);
duke@0 376 int words = _frame_slots >> shift;
duke@0 377 assert( words << shift == _frame_slots, "frame size must be properly aligned in LP64" );
duke@0 378 return words;
duke@0 379 }
duke@0 380
duke@0 381 // ============================================================================
duke@0 382 //------------------------------CompileWrapper---------------------------------
duke@0 383 class CompileWrapper : public StackObj {
duke@0 384 Compile *const _compile;
duke@0 385 public:
duke@0 386 CompileWrapper(Compile* compile);
duke@0 387
duke@0 388 ~CompileWrapper();
duke@0 389 };
duke@0 390
duke@0 391 CompileWrapper::CompileWrapper(Compile* compile) : _compile(compile) {
duke@0 392 // the Compile* pointer is stored in the current ciEnv:
duke@0 393 ciEnv* env = compile->env();
duke@0 394 assert(env == ciEnv::current(), "must already be a ciEnv active");
duke@0 395 assert(env->compiler_data() == NULL, "compile already active?");
duke@0 396 env->set_compiler_data(compile);
duke@0 397 assert(compile == Compile::current(), "sanity");
duke@0 398
duke@0 399 compile->set_type_dict(NULL);
duke@0 400 compile->set_type_hwm(NULL);
duke@0 401 compile->set_type_last_size(0);
duke@0 402 compile->set_last_tf(NULL, NULL);
duke@0 403 compile->set_indexSet_arena(NULL);
duke@0 404 compile->set_indexSet_free_block_list(NULL);
duke@0 405 compile->init_type_arena();
duke@0 406 Type::Initialize(compile);
duke@0 407 _compile->set_scratch_buffer_blob(NULL);
duke@0 408 _compile->begin_method();
duke@0 409 }
duke@0 410 CompileWrapper::~CompileWrapper() {
duke@0 411 _compile->end_method();
duke@0 412 if (_compile->scratch_buffer_blob() != NULL)
duke@0 413 BufferBlob::free(_compile->scratch_buffer_blob());
duke@0 414 _compile->env()->set_compiler_data(NULL);
duke@0 415 }
duke@0 416
duke@0 417
duke@0 418 //----------------------------print_compile_messages---------------------------
duke@0 419 void Compile::print_compile_messages() {
duke@0 420 #ifndef PRODUCT
duke@0 421 // Check if recompiling
duke@0 422 if (_subsume_loads == false && PrintOpto) {
duke@0 423 // Recompiling without allowing machine instructions to subsume loads
duke@0 424 tty->print_cr("*********************************************************");
duke@0 425 tty->print_cr("** Bailout: Recompile without subsuming loads **");
duke@0 426 tty->print_cr("*********************************************************");
duke@0 427 }
kvn@38 428 if (_do_escape_analysis != DoEscapeAnalysis && PrintOpto) {
kvn@38 429 // Recompiling without escape analysis
kvn@38 430 tty->print_cr("*********************************************************");
kvn@38 431 tty->print_cr("** Bailout: Recompile without escape analysis **");
kvn@38 432 tty->print_cr("*********************************************************");
kvn@38 433 }
duke@0 434 if (env()->break_at_compile()) {
twisti@605 435 // Open the debugger when compiling this method.
duke@0 436 tty->print("### Breaking when compiling: ");
duke@0 437 method()->print_short_name();
duke@0 438 tty->cr();
duke@0 439 BREAKPOINT;
duke@0 440 }
duke@0 441
duke@0 442 if( PrintOpto ) {
duke@0 443 if (is_osr_compilation()) {
duke@0 444 tty->print("[OSR]%3d", _compile_id);
duke@0 445 } else {
duke@0 446 tty->print("%3d", _compile_id);
duke@0 447 }
duke@0 448 }
duke@0 449 #endif
duke@0 450 }
duke@0 451
duke@0 452
kvn@1979 453 //-----------------------init_scratch_buffer_blob------------------------------
kvn@1979 454 // Construct a temporary BufferBlob and cache it for this compile.
twisti@1915 455 void Compile::init_scratch_buffer_blob(int const_size) {
kvn@1979 456 // If there is already a scratch buffer blob allocated and the
kvn@1979 457 // constant section is big enough, use it. Otherwise free the
kvn@1979 458 // current and allocate a new one.
kvn@1979 459 BufferBlob* blob = scratch_buffer_blob();
kvn@1979 460 if ((blob != NULL) && (const_size <= _scratch_const_size)) {
kvn@1979 461 // Use the current blob.
kvn@1979 462 } else {
kvn@1979 463 if (blob != NULL) {
kvn@1979 464 BufferBlob::free(blob);
kvn@1979 465 }
duke@0 466
kvn@1979 467 ResourceMark rm;
kvn@1979 468 _scratch_const_size = const_size;
kvn@1979 469 int size = (MAX_inst_size + MAX_stubs_size + _scratch_const_size);
kvn@1979 470 blob = BufferBlob::create("Compile::scratch_buffer", size);
kvn@1979 471 // Record the buffer blob for next time.
kvn@1979 472 set_scratch_buffer_blob(blob);
kvn@1979 473 // Have we run out of code space?
kvn@1979 474 if (scratch_buffer_blob() == NULL) {
kvn@1979 475 // Let CompilerBroker disable further compilations.
kvn@1979 476 record_failure("Not enough space for scratch buffer in CodeCache");
kvn@1979 477 return;
kvn@1979 478 }
kvn@163 479 }
duke@0 480
duke@0 481 // Initialize the relocation buffers
twisti@1668 482 relocInfo* locs_buf = (relocInfo*) blob->content_end() - MAX_locs_size;
duke@0 483 set_scratch_locs_memory(locs_buf);
duke@0 484 }
duke@0 485
duke@0 486
duke@0 487 //-----------------------scratch_emit_size-------------------------------------
duke@0 488 // Helper function that computes size by emitting code
duke@0 489 uint Compile::scratch_emit_size(const Node* n) {
twisti@1915 490 // Start scratch_emit_size section.
twisti@1915 491 set_in_scratch_emit_size(true);
twisti@1915 492
duke@0 493 // Emit into a trash buffer and count bytes emitted.
duke@0 494 // This is a pretty expensive way to compute a size,
duke@0 495 // but it works well enough if seldom used.
duke@0 496 // All common fixed-size instructions are given a size
duke@0 497 // method by the AD file.
duke@0 498 // Note that the scratch buffer blob and locs memory are
duke@0 499 // allocated at the beginning of the compile task, and
duke@0 500 // may be shared by several calls to scratch_emit_size.
duke@0 501 // The allocation of the scratch buffer blob is particularly
duke@0 502 // expensive, since it has to grab the code cache lock.
duke@0 503 BufferBlob* blob = this->scratch_buffer_blob();
duke@0 504 assert(blob != NULL, "Initialize BufferBlob at start");
duke@0 505 assert(blob->size() > MAX_inst_size, "sanity");
duke@0 506 relocInfo* locs_buf = scratch_locs_memory();
twisti@1668 507 address blob_begin = blob->content_begin();
duke@0 508 address blob_end = (address)locs_buf;
twisti@1668 509 assert(blob->content_contains(blob_end), "sanity");
duke@0 510 CodeBuffer buf(blob_begin, blob_end - blob_begin);
twisti@1915 511 buf.initialize_consts_size(_scratch_const_size);
duke@0 512 buf.initialize_stubs_size(MAX_stubs_size);
duke@0 513 assert(locs_buf != NULL, "sanity");
twisti@1915 514 int lsize = MAX_locs_size / 3;
twisti@1915 515 buf.consts()->initialize_shared_locs(&locs_buf[lsize * 0], lsize);
twisti@1915 516 buf.insts()->initialize_shared_locs( &locs_buf[lsize * 1], lsize);
twisti@1915 517 buf.stubs()->initialize_shared_locs( &locs_buf[lsize * 2], lsize);
twisti@1915 518
twisti@1915 519 // Do the emission.
kvn@2664 520
kvn@2664 521 Label fakeL; // Fake label for branch instructions.
kvn@2678 522 Label* saveL = NULL;
kvn@2678 523 uint save_bnum = 0;
kvn@2678 524 bool is_branch = n->is_MachBranch();
kvn@2664 525 if (is_branch) {
kvn@2664 526 MacroAssembler masm(&buf);
kvn@2664 527 masm.bind(fakeL);
kvn@2678 528 n->as_MachBranch()->save_label(&saveL, &save_bnum);
kvn@2678 529 n->as_MachBranch()->label_set(&fakeL, 0);
kvn@2664 530 }
duke@0 531 n->emit(buf, this->regalloc());
kvn@2678 532 if (is_branch) // Restore label.
kvn@2678 533 n->as_MachBranch()->label_set(saveL, save_bnum);
twisti@1915 534
twisti@1915 535 // End scratch_emit_size section.
twisti@1915 536 set_in_scratch_emit_size(false);
twisti@1915 537
twisti@1668 538 return buf.insts_size();
duke@0 539 }
duke@0 540
duke@0 541
duke@0 542 // ============================================================================
duke@0 543 //------------------------------Compile standard-------------------------------
duke@0 544 debug_only( int Compile::_debug_idx = 100000; )
duke@0 545
duke@0 546 // Compile a method. entry_bci is -1 for normal compilations and indicates
duke@0 547 // the continuation bci for on stack replacement.
duke@0 548
duke@0 549
kvn@38 550 Compile::Compile( ciEnv* ci_env, C2Compiler* compiler, ciMethod* target, int osr_bci, bool subsume_loads, bool do_escape_analysis )
duke@0 551 : Phase(Compiler),
duke@0 552 _env(ci_env),
duke@0 553 _log(ci_env->log()),
duke@0 554 _compile_id(ci_env->compile_id()),
duke@0 555 _save_argument_registers(false),
duke@0 556 _stub_name(NULL),
duke@0 557 _stub_function(NULL),
duke@0 558 _stub_entry_point(NULL),
duke@0 559 _method(target),
duke@0 560 _entry_bci(osr_bci),
duke@0 561 _initial_gvn(NULL),
duke@0 562 _for_igvn(NULL),
duke@0 563 _warm_calls(NULL),
duke@0 564 _subsume_loads(subsume_loads),
kvn@38 565 _do_escape_analysis(do_escape_analysis),
duke@0 566 _failure_reason(NULL),
duke@0 567 _code_buffer("Compile::Fill_buffer"),
duke@0 568 _orig_pc_slot(0),
duke@0 569 _orig_pc_slot_offset_in_bytes(0),
twisti@1265 570 _has_method_handle_invokes(false),
twisti@1915 571 _mach_constant_base_node(NULL),
duke@0 572 _node_bundling_limit(0),
duke@0 573 _node_bundling_base(NULL),
kvn@859 574 _java_calls(0),
kvn@859 575 _inner_loops(0),
twisti@1915 576 _scratch_const_size(-1),
twisti@1915 577 _in_scratch_emit_size(false),
duke@0 578 #ifndef PRODUCT
duke@0 579 _trace_opto_output(TraceOptoOutput || method()->has_option("TraceOptoOutput")),
duke@0 580 _printer(IdealGraphPrinter::printer()),
duke@0 581 #endif
duke@0 582 _congraph(NULL) {
duke@0 583 C = this;
duke@0 584
duke@0 585 CompileWrapper cw(this);
duke@0 586 #ifndef PRODUCT
duke@0 587 if (TimeCompiler2) {
duke@0 588 tty->print(" ");
duke@0 589 target->holder()->name()->print();
duke@0 590 tty->print(".");
duke@0 591 target->print_short_name();
duke@0 592 tty->print(" ");
duke@0 593 }
duke@0 594 TraceTime t1("Total compilation time", &_t_totalCompilation, TimeCompiler, TimeCompiler2);
duke@0 595 TraceTime t2(NULL, &_t_methodCompilation, TimeCompiler, false);
jrose@100 596 bool print_opto_assembly = PrintOptoAssembly || _method->has_option("PrintOptoAssembly");
jrose@100 597 if (!print_opto_assembly) {
jrose@100 598 bool print_assembly = (PrintAssembly || _method->should_print_assembly());
jrose@100 599 if (print_assembly && !Disassembler::can_decode()) {
jrose@100 600 tty->print_cr("PrintAssembly request changed to PrintOptoAssembly");
jrose@100 601 print_opto_assembly = true;
jrose@100 602 }
jrose@100 603 }
jrose@100 604 set_print_assembly(print_opto_assembly);
never@367 605 set_parsed_irreducible_loop(false);
duke@0 606 #endif
duke@0 607
duke@0 608 if (ProfileTraps) {
duke@0 609 // Make sure the method being compiled gets its own MDO,
duke@0 610 // so we can at least track the decompile_count().
iveresov@1914 611 method()->ensure_method_data();
duke@0 612 }
duke@0 613
duke@0 614 Init(::AliasLevel);
duke@0 615
duke@0 616
duke@0 617 print_compile_messages();
duke@0 618
duke@0 619 if (UseOldInlining || PrintCompilation NOT_PRODUCT( || PrintOpto) )
duke@0 620 _ilt = InlineTree::build_inline_tree_root();
duke@0 621 else
duke@0 622 _ilt = NULL;
duke@0 623
duke@0 624 // Even if NO memory addresses are used, MergeMem nodes must have at least 1 slice
duke@0 625 assert(num_alias_types() >= AliasIdxRaw, "");
duke@0 626
duke@0 627 #define MINIMUM_NODE_HASH 1023
duke@0 628 // Node list that Iterative GVN will start with
duke@0 629 Unique_Node_List for_igvn(comp_arena());
duke@0 630 set_for_igvn(&for_igvn);
duke@0 631
duke@0 632 // GVN that will be run immediately on new nodes
duke@0 633 uint estimated_size = method()->code_size()*4+64;
duke@0 634 estimated_size = (estimated_size < MINIMUM_NODE_HASH ? MINIMUM_NODE_HASH : estimated_size);
duke@0 635 PhaseGVN gvn(node_arena(), estimated_size);
duke@0 636 set_initial_gvn(&gvn);
duke@0 637
duke@0 638 { // Scope for timing the parser
duke@0 639 TracePhase t3("parse", &_t_parser, true);
duke@0 640
duke@0 641 // Put top into the hash table ASAP.
duke@0 642 initial_gvn()->transform_no_reclaim(top());
duke@0 643
duke@0 644 // Set up tf(), start(), and find a CallGenerator.
johnc@2346 645 CallGenerator* cg = NULL;
duke@0 646 if (is_osr_compilation()) {
duke@0 647 const TypeTuple *domain = StartOSRNode::osr_domain();
duke@0 648 const TypeTuple *range = TypeTuple::make_range(method()->signature());
duke@0 649 init_tf(TypeFunc::make(domain, range));
duke@0 650 StartNode* s = new (this, 2) StartOSRNode(root(), domain);
duke@0 651 initial_gvn()->set_type_bottom(s);
duke@0 652 init_start(s);
duke@0 653 cg = CallGenerator::for_osr(method(), entry_bci());
duke@0 654 } else {
duke@0 655 // Normal case.
duke@0 656 init_tf(TypeFunc::make(method()));
duke@0 657 StartNode* s = new (this, 2) StartNode(root(), tf()->domain());
duke@0 658 initial_gvn()->set_type_bottom(s);
duke@0 659 init_start(s);
johnc@2346 660 if (method()->intrinsic_id() == vmIntrinsics::_Reference_get && UseG1GC) {
johnc@2346 661 // With java.lang.ref.reference.get() we must go through the
johnc@2346 662 // intrinsic when G1 is enabled - even when get() is the root
johnc@2346 663 // method of the compile - so that, if necessary, the value in
johnc@2346 664 // the referent field of the reference object gets recorded by
johnc@2346 665 // the pre-barrier code.
johnc@2346 666 // Specifically, if G1 is enabled, the value in the referent
johnc@2346 667 // field is recorded by the G1 SATB pre barrier. This will
johnc@2346 668 // result in the referent being marked live and the reference
johnc@2346 669 // object removed from the list of discovered references during
johnc@2346 670 // reference processing.
johnc@2346 671 cg = find_intrinsic(method(), false);
johnc@2346 672 }
johnc@2346 673 if (cg == NULL) {
johnc@2346 674 float past_uses = method()->interpreter_invocation_count();
johnc@2346 675 float expected_uses = past_uses;
johnc@2346 676 cg = CallGenerator::for_inline(method(), expected_uses);
johnc@2346 677 }
duke@0 678 }
duke@0 679 if (failing()) return;
duke@0 680 if (cg == NULL) {
duke@0 681 record_method_not_compilable_all_tiers("cannot parse method");
duke@0 682 return;
duke@0 683 }
duke@0 684 JVMState* jvms = build_start_state(start(), tf());
duke@0 685 if ((jvms = cg->generate(jvms)) == NULL) {
duke@0 686 record_method_not_compilable("method parse failed");
duke@0 687 return;
duke@0 688 }
duke@0 689 GraphKit kit(jvms);
duke@0 690
duke@0 691 if (!kit.stopped()) {
duke@0 692 // Accept return values, and transfer control we know not where.
duke@0 693 // This is done by a special, unique ReturnNode bound to root.
duke@0 694 return_values(kit.jvms());
duke@0 695 }
duke@0 696
duke@0 697 if (kit.has_exceptions()) {
duke@0 698 // Any exceptions that escape from this call must be rethrown
duke@0 699 // to whatever caller is dynamically above us on the stack.
duke@0 700 // This is done by a special, unique RethrowNode bound to root.
duke@0 701 rethrow_exceptions(kit.transfer_exceptions_into_jvms());
duke@0 702 }
duke@0 703
never@1080 704 if (!failing() && has_stringbuilder()) {
never@1080 705 {
never@1080 706 // remove useless nodes to make the usage analysis simpler
never@1080 707 ResourceMark rm;
never@1080 708 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
never@1080 709 }
never@1080 710
never@1080 711 {
never@1080 712 ResourceMark rm;
never@1080 713 print_method("Before StringOpts", 3);
never@1080 714 PhaseStringOpts pso(initial_gvn(), &for_igvn);
never@1080 715 print_method("After StringOpts", 3);
never@1080 716 }
never@1080 717
never@1080 718 // now inline anything that we skipped the first time around
never@1080 719 while (_late_inlines.length() > 0) {
never@1080 720 CallGenerator* cg = _late_inlines.pop();
never@1080 721 cg->do_late_inline();
never@1080 722 }
never@1080 723 }
never@1080 724 assert(_late_inlines.length() == 0, "should have been processed");
never@1080 725
never@417 726 print_method("Before RemoveUseless", 3);
never@367 727
duke@0 728 // Remove clutter produced by parsing.
duke@0 729 if (!failing()) {
duke@0 730 ResourceMark rm;
duke@0 731 PhaseRemoveUseless pru(initial_gvn(), &for_igvn);
duke@0 732 }
duke@0 733 }
duke@0 734
duke@0 735 // Note: Large methods are capped off in do_one_bytecode().
duke@0 736 if (failing()) return;
duke@0 737
duke@0 738 // After parsing, node notes are no longer automagic.
duke@0 739 // They must be propagated by register_new_node_with_optimizer(),
duke@0 740 // clone(), or the like.
duke@0 741 set_default_node_notes(NULL);
duke@0 742
duke@0 743 for (;;) {
duke@0 744 int successes = Inline_Warm();
duke@0 745 if (failing()) return;
duke@0 746 if (successes == 0) break;
duke@0 747 }
duke@0 748
duke@0 749 // Drain the list.
duke@0 750 Finish_Warm();
duke@0 751 #ifndef PRODUCT
duke@0 752 if (_printer) {
duke@0 753 _printer->print_inlining(this);
duke@0 754 }
duke@0 755 #endif
duke@0 756
duke@0 757 if (failing()) return;
duke@0 758 NOT_PRODUCT( verify_graph_edges(); )
duke@0 759
duke@0 760 // Now optimize
duke@0 761 Optimize();
duke@0 762 if (failing()) return;
duke@0 763 NOT_PRODUCT( verify_graph_edges(); )
duke@0 764
duke@0 765 #ifndef PRODUCT
duke@0 766 if (PrintIdeal) {
duke@0 767 ttyLocker ttyl; // keep the following output all in one block
duke@0 768 // This output goes directly to the tty, not the compiler log.
duke@0 769 // To enable tools to match it up with the compilation activity,
duke@0 770 // be sure to tag this tty output with the compile ID.
duke@0 771 if (xtty != NULL) {
duke@0 772 xtty->head("ideal compile_id='%d'%s", compile_id(),
duke@0 773 is_osr_compilation() ? " compile_kind='osr'" :
duke@0 774 "");
duke@0 775 }
duke@0 776 root()->dump(9999);
duke@0 777 if (xtty != NULL) {
duke@0 778 xtty->tail("ideal");
duke@0 779 }
duke@0 780 }
duke@0 781 #endif
duke@0 782
duke@0 783 // Now that we know the size of all the monitors we can add a fixed slot
duke@0 784 // for the original deopt pc.
duke@0 785
duke@0 786 _orig_pc_slot = fixed_slots();
duke@0 787 int next_slot = _orig_pc_slot + (sizeof(address) / VMRegImpl::stack_slot_size);
duke@0 788 set_fixed_slots(next_slot);
duke@0 789
duke@0 790 // Now generate code
duke@0 791 Code_Gen();
duke@0 792 if (failing()) return;
duke@0 793
duke@0 794 // Check if we want to skip execution of all compiled code.
duke@0 795 {
duke@0 796 #ifndef PRODUCT
duke@0 797 if (OptoNoExecute) {
duke@0 798 record_method_not_compilable("+OptoNoExecute"); // Flag as failed
duke@0 799 return;
duke@0 800 }
duke@0 801 TracePhase t2("install_code", &_t_registerMethod, TimeCompiler);
duke@0 802 #endif
duke@0 803
duke@0 804 if (is_osr_compilation()) {
duke@0 805 _code_offsets.set_value(CodeOffsets::Verified_Entry, 0);
duke@0 806 _code_offsets.set_value(CodeOffsets::OSR_Entry, _first_block_size);
duke@0 807 } else {
duke@0 808 _code_offsets.set_value(CodeOffsets::Verified_Entry, _first_block_size);
duke@0 809 _code_offsets.set_value(CodeOffsets::OSR_Entry, 0);
duke@0 810 }
duke@0 811
duke@0 812 env()->register_method(_method, _entry_bci,
duke@0 813 &_code_offsets,
duke@0 814 _orig_pc_slot_offset_in_bytes,
duke@0 815 code_buffer(),
duke@0 816 frame_size_in_words(), _oop_map_set,
duke@0 817 &_handler_table, &_inc_table,
duke@0 818 compiler,
duke@0 819 env()->comp_level(),
duke@0 820 has_unsafe_access()
duke@0 821 );
duke@0 822 }
duke@0 823 }
duke@0 824
duke@0 825 //------------------------------Compile----------------------------------------
duke@0 826 // Compile a runtime stub
duke@0 827 Compile::Compile( ciEnv* ci_env,
duke@0 828 TypeFunc_generator generator,
duke@0 829 address stub_function,
duke@0 830 const char *stub_name,
duke@0 831 int is_fancy_jump,
duke@0 832 bool pass_tls,
duke@0 833 bool save_arg_registers,
duke@0 834 bool return_pc )
duke@0 835 : Phase(Compiler),
duke@0 836 _env(ci_env),
duke@0 837 _log(ci_env->log()),
duke@0 838 _compile_id(-1),
duke@0 839 _save_argument_registers(save_arg_registers),
duke@0 840 _method(NULL),
duke@0 841 _stub_name(stub_name),
duke@0 842 _stub_function(stub_function),
duke@0 843 _stub_entry_point(NULL),
duke@0 844 _entry_bci(InvocationEntryBci),
duke@0 845 _initial_gvn(NULL),
duke@0 846 _for_igvn(NULL),
duke@0 847 _warm_calls(NULL),
duke@0 848 _orig_pc_slot(0),
duke@0 849 _orig_pc_slot_offset_in_bytes(0),
duke@0 850 _subsume_loads(true),
kvn@38 851 _do_escape_analysis(false),
duke@0 852 _failure_reason(NULL),
duke@0 853 _code_buffer("Compile::Fill_buffer"),
twisti@1265 854 _has_method_handle_invokes(false),
twisti@1915 855 _mach_constant_base_node(NULL),
duke@0 856 _node_bundling_limit(0),
duke@0 857 _node_bundling_base(NULL),
kvn@859 858 _java_calls(0),
kvn@859 859 _inner_loops(0),
duke@0 860 #ifndef PRODUCT
duke@0 861 _trace_opto_output(TraceOptoOutput),
duke@0 862 _printer(NULL),
duke@0 863 #endif
duke@0 864 _congraph(NULL) {
duke@0 865 C = this;
duke@0 866
duke@0 867 #ifndef PRODUCT
duke@0 868 TraceTime t1(NULL, &_t_totalCompilation, TimeCompiler, false);
duke@0 869 TraceTime t2(NULL, &_t_stubCompilation, TimeCompiler, false);
duke@0 870 set_print_assembly(PrintFrameConverterAssembly);
never@367 871 set_parsed_irreducible_loop(false);
duke@0 872 #endif
duke@0 873 CompileWrapper cw(this);
duke@0 874 Init(/*AliasLevel=*/ 0);
duke@0 875 init_tf((*generator)());
duke@0 876
duke@0 877 {
duke@0 878 // The following is a dummy for the sake of GraphKit::gen_stub
duke@0 879 Unique_Node_List for_igvn(comp_arena());
duke@0 880 set_for_igvn(&for_igvn); // not used, but some GraphKit guys push on this
duke@0 881 PhaseGVN gvn(Thread::current()->resource_area(),255);
duke@0 882 set_initial_gvn(&gvn); // not significant, but GraphKit guys use it pervasively
duke@0 883 gvn.transform_no_reclaim(top());
duke@0 884
duke@0 885 GraphKit kit;
duke@0 886 kit.gen_stub(stub_function, stub_name, is_fancy_jump, pass_tls, return_pc);
duke@0 887 }
duke@0 888
duke@0 889 NOT_PRODUCT( verify_graph_edges(); )
duke@0 890 Code_Gen();
duke@0 891 if (failing()) return;
duke@0 892
duke@0 893
duke@0 894 // Entry point will be accessed using compile->stub_entry_point();
duke@0 895 if (code_buffer() == NULL) {
duke@0 896 Matcher::soft_match_failure();
duke@0 897 } else {
duke@0 898 if (PrintAssembly && (WizardMode || Verbose))
duke@0 899 tty->print_cr("### Stub::%s", stub_name);
duke@0 900
duke@0 901 if (!failing()) {
duke@0 902 assert(_fixed_slots == 0, "no fixed slots used for runtime stubs");
duke@0 903
duke@0 904 // Make the NMethod
duke@0 905 // For now we mark the frame as never safe for profile stackwalking
duke@0 906 RuntimeStub *rs = RuntimeStub::new_runtime_stub(stub_name,
duke@0 907 code_buffer(),
duke@0 908 CodeOffsets::frame_never_safe,
duke@0 909 // _code_offsets.value(CodeOffsets::Frame_Complete),
duke@0 910 frame_size_in_words(),
duke@0 911 _oop_map_set,
duke@0 912 save_arg_registers);
duke@0 913 assert(rs != NULL && rs->is_runtime_stub(), "sanity check");
duke@0 914
duke@0 915 _stub_entry_point = rs->entry_point();
duke@0 916 }
duke@0 917 }
duke@0 918 }
duke@0 919
duke@0 920 #ifndef PRODUCT
duke@0 921 void print_opto_verbose_signature( const TypeFunc *j_sig, const char *stub_name ) {
duke@0 922 if(PrintOpto && Verbose) {
duke@0 923 tty->print("%s ", stub_name); j_sig->print_flattened(); tty->cr();
duke@0 924 }
duke@0 925 }
duke@0 926 #endif
duke@0 927
duke@0 928 void Compile::print_codes() {
duke@0 929 }
duke@0 930
duke@0 931 //------------------------------Init-------------------------------------------
duke@0 932 // Prepare for a single compilation
duke@0 933 void Compile::Init(int aliaslevel) {
duke@0 934 _unique = 0;
duke@0 935 _regalloc = NULL;
duke@0 936
duke@0 937 _tf = NULL; // filled in later
duke@0 938 _top = NULL; // cached later
duke@0 939 _matcher = NULL; // filled in later
duke@0 940 _cfg = NULL; // filled in later
duke@0 941
duke@0 942 set_24_bit_selection_and_mode(Use24BitFP, false);
duke@0 943
duke@0 944 _node_note_array = NULL;
duke@0 945 _default_node_notes = NULL;
duke@0 946
duke@0 947 _immutable_memory = NULL; // filled in at first inquiry
duke@0 948
duke@0 949 // Globally visible Nodes
duke@0 950 // First set TOP to NULL to give safe behavior during creation of RootNode
duke@0 951 set_cached_top_node(NULL);
duke@0 952 set_root(new (this, 3) RootNode());
duke@0 953 // Now that you have a Root to point to, create the real TOP
duke@0 954 set_cached_top_node( new (this, 1) ConNode(Type::TOP) );
duke@0 955 set_recent_alloc(NULL, NULL);
duke@0 956
duke@0 957 // Create Debug Information Recorder to record scopes, oopmaps, etc.
duke@0 958 env()->set_oop_recorder(new OopRecorder(comp_arena()));
duke@0 959 env()->set_debug_info(new DebugInformationRecorder(env()->oop_recorder()));
duke@0 960 env()->set_dependencies(new Dependencies(env()));
duke@0 961
duke@0 962 _fixed_slots = 0;
duke@0 963 set_has_split_ifs(false);
duke@0 964 set_has_loops(has_method() && method()->has_loops()); // first approximation
never@1080 965 set_has_stringbuilder(false);
duke@0 966 _trap_can_recompile = false; // no traps emitted yet
duke@0 967 _major_progress = true; // start out assuming good things will happen
duke@0 968 set_has_unsafe_access(false);
duke@0 969 Copy::zero_to_bytes(_trap_hist, sizeof(_trap_hist));
duke@0 970 set_decompile_count(0);
duke@0 971
rasbold@418 972 set_do_freq_based_layout(BlockLayoutByFrequency || method_has_option("BlockLayoutByFrequency"));
iveresov@1703 973 set_num_loop_opts(LoopOptsCount);
iveresov@1703 974 set_do_inlining(Inline);
iveresov@1703 975 set_max_inline_size(MaxInlineSize);
iveresov@1703 976 set_freq_inline_size(FreqInlineSize);
iveresov@1703 977 set_do_scheduling(OptoScheduling);
iveresov@1703 978 set_do_count_invocations(false);
iveresov@1703 979 set_do_method_data_update(false);
duke@0 980
duke@0 981 if (debug_info()->recording_non_safepoints()) {
duke@0 982 set_node_note_array(new(comp_arena()) GrowableArray<Node_Notes*>
duke@0 983 (comp_arena(), 8, 0, NULL));
duke@0 984 set_default_node_notes(Node_Notes::make(this));
duke@0 985 }
duke@0 986
duke@0 987 // // -- Initialize types before each compile --
duke@0 988 // // Update cached type information
duke@0 989 // if( _method && _method->constants() )
duke@0 990 // Type::update_loaded_types(_method, _method->constants());
duke@0 991
duke@0 992 // Init alias_type map.
kvn@38 993 if (!_do_escape_analysis && aliaslevel == 3)
duke@0 994 aliaslevel = 2; // No unique types without escape analysis
duke@0 995 _AliasLevel = aliaslevel;
duke@0 996 const int grow_ats = 16;
duke@0 997 _max_alias_types = grow_ats;
duke@0 998 _alias_types = NEW_ARENA_ARRAY(comp_arena(), AliasType*, grow_ats);
duke@0 999 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, grow_ats);
duke@0 1000 Copy::zero_to_bytes(ats, sizeof(AliasType)*grow_ats);
duke@0 1001 {
duke@0 1002 for (int i = 0; i < grow_ats; i++) _alias_types[i] = &ats[i];
duke@0 1003 }
duke@0 1004 // Initialize the first few types.
duke@0 1005 _alias_types[AliasIdxTop]->Init(AliasIdxTop, NULL);
duke@0 1006 _alias_types[AliasIdxBot]->Init(AliasIdxBot, TypePtr::BOTTOM);
duke@0 1007 _alias_types[AliasIdxRaw]->Init(AliasIdxRaw, TypeRawPtr::BOTTOM);
duke@0 1008 _num_alias_types = AliasIdxRaw+1;
duke@0 1009 // Zero out the alias type cache.
duke@0 1010 Copy::zero_to_bytes(_alias_cache, sizeof(_alias_cache));
duke@0 1011 // A NULL adr_type hits in the cache right away. Preload the right answer.
duke@0 1012 probe_alias_cache(NULL)->_index = AliasIdxTop;
duke@0 1013
duke@0 1014 _intrinsics = NULL;
kvn@1605 1015 _macro_nodes = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
kvn@1605 1016 _predicate_opaqs = new(comp_arena()) GrowableArray<Node*>(comp_arena(), 8, 0, NULL);
duke@0 1017 register_library_intrinsics();
duke@0 1018 }
duke@0 1019
duke@0 1020 //---------------------------init_start----------------------------------------
duke@0 1021 // Install the StartNode on this compile object.
duke@0 1022 void Compile::init_start(StartNode* s) {
duke@0 1023 if (failing())
duke@0 1024 return; // already failing
duke@0 1025 assert(s == start(), "");
duke@0 1026 }
duke@0 1027
duke@0 1028 StartNode* Compile::start() const {
duke@0 1029 assert(!failing(), "");
duke@0 1030 for (DUIterator_Fast imax, i = root()->fast_outs(imax); i < imax; i++) {
duke@0 1031 Node* start = root()->fast_out(i);
duke@0 1032 if( start->is_Start() )
duke@0 1033 return start->as_Start();
duke@0 1034 }
duke@0 1035 ShouldNotReachHere();
duke@0 1036 return NULL;
duke@0 1037 }
duke@0 1038
duke@0 1039 //-------------------------------immutable_memory-------------------------------------
duke@0 1040 // Access immutable memory
duke@0 1041 Node* Compile::immutable_memory() {
duke@0 1042 if (_immutable_memory != NULL) {
duke@0 1043 return _immutable_memory;
duke@0 1044 }
duke@0 1045 StartNode* s = start();
duke@0 1046 for (DUIterator_Fast imax, i = s->fast_outs(imax); true; i++) {
duke@0 1047 Node *p = s->fast_out(i);
duke@0 1048 if (p != s && p->as_Proj()->_con == TypeFunc::Memory) {
duke@0 1049 _immutable_memory = p;
duke@0 1050 return _immutable_memory;
duke@0 1051 }
duke@0 1052 }
duke@0 1053 ShouldNotReachHere();
duke@0 1054 return NULL;
duke@0 1055 }
duke@0 1056
duke@0 1057 //----------------------set_cached_top_node------------------------------------
duke@0 1058 // Install the cached top node, and make sure Node::is_top works correctly.
duke@0 1059 void Compile::set_cached_top_node(Node* tn) {
duke@0 1060 if (tn != NULL) verify_top(tn);
duke@0 1061 Node* old_top = _top;
duke@0 1062 _top = tn;
duke@0 1063 // Calling Node::setup_is_top allows the nodes the chance to adjust
duke@0 1064 // their _out arrays.
duke@0 1065 if (_top != NULL) _top->setup_is_top();
duke@0 1066 if (old_top != NULL) old_top->setup_is_top();
duke@0 1067 assert(_top == NULL || top()->is_top(), "");
duke@0 1068 }
duke@0 1069
duke@0 1070 #ifndef PRODUCT
duke@0 1071 void Compile::verify_top(Node* tn) const {
duke@0 1072 if (tn != NULL) {
duke@0 1073 assert(tn->is_Con(), "top node must be a constant");
duke@0 1074 assert(((ConNode*)tn)->type() == Type::TOP, "top node must have correct type");
duke@0 1075 assert(tn->in(0) != NULL, "must have live top node");
duke@0 1076 }
duke@0 1077 }
duke@0 1078 #endif
duke@0 1079
duke@0 1080
duke@0 1081 ///-------------------Managing Per-Node Debug & Profile Info-------------------
duke@0 1082
duke@0 1083 void Compile::grow_node_notes(GrowableArray<Node_Notes*>* arr, int grow_by) {
duke@0 1084 guarantee(arr != NULL, "");
duke@0 1085 int num_blocks = arr->length();
duke@0 1086 if (grow_by < num_blocks) grow_by = num_blocks;
duke@0 1087 int num_notes = grow_by * _node_notes_block_size;
duke@0 1088 Node_Notes* notes = NEW_ARENA_ARRAY(node_arena(), Node_Notes, num_notes);
duke@0 1089 Copy::zero_to_bytes(notes, num_notes * sizeof(Node_Notes));
duke@0 1090 while (num_notes > 0) {
duke@0 1091 arr->append(notes);
duke@0 1092 notes += _node_notes_block_size;
duke@0 1093 num_notes -= _node_notes_block_size;
duke@0 1094 }
duke@0 1095 assert(num_notes == 0, "exact multiple, please");
duke@0 1096 }
duke@0 1097
duke@0 1098 bool Compile::copy_node_notes_to(Node* dest, Node* source) {
duke@0 1099 if (source == NULL || dest == NULL) return false;
duke@0 1100
duke@0 1101 if (dest->is_Con())
duke@0 1102 return false; // Do not push debug info onto constants.
duke@0 1103
duke@0 1104 #ifdef ASSERT
duke@0 1105 // Leave a bread crumb trail pointing to the original node:
duke@0 1106 if (dest != NULL && dest != source && dest->debug_orig() == NULL) {
duke@0 1107 dest->set_debug_orig(source);
duke@0 1108 }
duke@0 1109 #endif
duke@0 1110
duke@0 1111 if (node_note_array() == NULL)
duke@0 1112 return false; // Not collecting any notes now.
duke@0 1113
duke@0 1114 // This is a copy onto a pre-existing node, which may already have notes.
duke@0 1115 // If both nodes have notes, do not overwrite any pre-existing notes.
duke@0 1116 Node_Notes* source_notes = node_notes_at(source->_idx);
duke@0 1117 if (source_notes == NULL || source_notes->is_clear()) return false;
duke@0 1118 Node_Notes* dest_notes = node_notes_at(dest->_idx);
duke@0 1119 if (dest_notes == NULL || dest_notes->is_clear()) {
duke@0 1120 return set_node_notes_at(dest->_idx, source_notes);
duke@0 1121 }
duke@0 1122
duke@0 1123 Node_Notes merged_notes = (*source_notes);
duke@0 1124 // The order of operations here ensures that dest notes will win...
duke@0 1125 merged_notes.update_from(dest_notes);
duke@0 1126 return set_node_notes_at(dest->_idx, &merged_notes);
duke@0 1127 }
duke@0 1128
duke@0 1129
duke@0 1130 //--------------------------allow_range_check_smearing-------------------------
duke@0 1131 // Gating condition for coalescing similar range checks.
duke@0 1132 // Sometimes we try 'speculatively' replacing a series of a range checks by a
duke@0 1133 // single covering check that is at least as strong as any of them.
duke@0 1134 // If the optimization succeeds, the simplified (strengthened) range check
duke@0 1135 // will always succeed. If it fails, we will deopt, and then give up
duke@0 1136 // on the optimization.
duke@0 1137 bool Compile::allow_range_check_smearing() const {
duke@0 1138 // If this method has already thrown a range-check,
duke@0 1139 // assume it was because we already tried range smearing
duke@0 1140 // and it failed.
duke@0 1141 uint already_trapped = trap_count(Deoptimization::Reason_range_check);
duke@0 1142 return !already_trapped;
duke@0 1143 }
duke@0 1144
duke@0 1145
duke@0 1146 //------------------------------flatten_alias_type-----------------------------
duke@0 1147 const TypePtr *Compile::flatten_alias_type( const TypePtr *tj ) const {
duke@0 1148 int offset = tj->offset();
duke@0 1149 TypePtr::PTR ptr = tj->ptr();
duke@0 1150
kvn@247 1151 // Known instance (scalarizable allocation) alias only with itself.
kvn@247 1152 bool is_known_inst = tj->isa_oopptr() != NULL &&
kvn@247 1153 tj->is_oopptr()->is_known_instance();
kvn@247 1154
duke@0 1155 // Process weird unsafe references.
duke@0 1156 if (offset == Type::OffsetBot && (tj->isa_instptr() /*|| tj->isa_klassptr()*/)) {
duke@0 1157 assert(InlineUnsafeOps, "indeterminate pointers come only from unsafe ops");
kvn@247 1158 assert(!is_known_inst, "scalarizable allocation should not have unsafe references");
duke@0 1159 tj = TypeOopPtr::BOTTOM;
duke@0 1160 ptr = tj->ptr();
duke@0 1161 offset = tj->offset();
duke@0 1162 }
duke@0 1163
duke@0 1164 // Array pointers need some flattening
duke@0 1165 const TypeAryPtr *ta = tj->isa_aryptr();
kvn@247 1166 if( ta && is_known_inst ) {
kvn@247 1167 if ( offset != Type::OffsetBot &&
kvn@247 1168 offset > arrayOopDesc::length_offset_in_bytes() ) {
kvn@247 1169 offset = Type::OffsetBot; // Flatten constant access into array body only
kvn@247 1170 tj = ta = TypeAryPtr::make(ptr, ta->ary(), ta->klass(), true, offset, ta->instance_id());
kvn@247 1171 }
kvn@247 1172 } else if( ta && _AliasLevel >= 2 ) {
duke@0 1173 // For arrays indexed by constant indices, we flatten the alias
duke@0 1174 // space to include all of the array body. Only the header, klass
duke@0 1175 // and array length can be accessed un-aliased.
duke@0 1176 if( offset != Type::OffsetBot ) {
duke@0 1177 if( ta->const_oop() ) { // methodDataOop or methodOop
duke@0 1178 offset = Type::OffsetBot; // Flatten constant access into array body
kvn@247 1179 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),ta->ary(),ta->klass(),false,offset);
duke@0 1180 } else if( offset == arrayOopDesc::length_offset_in_bytes() ) {
duke@0 1181 // range is OK as-is.
duke@0 1182 tj = ta = TypeAryPtr::RANGE;
duke@0 1183 } else if( offset == oopDesc::klass_offset_in_bytes() ) {
duke@0 1184 tj = TypeInstPtr::KLASS; // all klass loads look alike
duke@0 1185 ta = TypeAryPtr::RANGE; // generic ignored junk
duke@0 1186 ptr = TypePtr::BotPTR;
duke@0 1187 } else if( offset == oopDesc::mark_offset_in_bytes() ) {
duke@0 1188 tj = TypeInstPtr::MARK;
duke@0 1189 ta = TypeAryPtr::RANGE; // generic ignored junk
duke@0 1190 ptr = TypePtr::BotPTR;
duke@0 1191 } else { // Random constant offset into array body
duke@0 1192 offset = Type::OffsetBot; // Flatten constant access into array body
kvn@247 1193 tj = ta = TypeAryPtr::make(ptr,ta->ary(),ta->klass(),false,offset);
duke@0 1194 }
duke@0 1195 }
duke@0 1196 // Arrays of fixed size alias with arrays of unknown size.
duke@0 1197 if (ta->size() != TypeInt::POS) {
duke@0 1198 const TypeAry *tary = TypeAry::make(ta->elem(), TypeInt::POS);
kvn@247 1199 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,ta->klass(),false,offset);
duke@0 1200 }
duke@0 1201 // Arrays of known objects become arrays of unknown objects.
coleenp@113 1202 if (ta->elem()->isa_narrowoop() && ta->elem() != TypeNarrowOop::BOTTOM) {
coleenp@113 1203 const TypeAry *tary = TypeAry::make(TypeNarrowOop::BOTTOM, ta->size());
kvn@247 1204 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
coleenp@113 1205 }
duke@0 1206 if (ta->elem()->isa_oopptr() && ta->elem() != TypeInstPtr::BOTTOM) {
duke@0 1207 const TypeAry *tary = TypeAry::make(TypeInstPtr::BOTTOM, ta->size());
kvn@247 1208 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,NULL,false,offset);
duke@0 1209 }
duke@0 1210 // Arrays of bytes and of booleans both use 'bastore' and 'baload' so
duke@0 1211 // cannot be distinguished by bytecode alone.
duke@0 1212 if (ta->elem() == TypeInt::BOOL) {
duke@0 1213 const TypeAry *tary = TypeAry::make(TypeInt::BYTE, ta->size());
duke@0 1214 ciKlass* aklass = ciTypeArrayKlass::make(T_BYTE);
kvn@247 1215 tj = ta = TypeAryPtr::make(ptr,ta->const_oop(),tary,aklass,false,offset);
duke@0 1216 }
duke@0 1217 // During the 2nd round of IterGVN, NotNull castings are removed.
duke@0 1218 // Make sure the Bottom and NotNull variants alias the same.
duke@0 1219 // Also, make sure exact and non-exact variants alias the same.
duke@0 1220 if( ptr == TypePtr::NotNull || ta->klass_is_exact() ) {
kvn@2614 1221 tj = ta = TypeAryPtr::make(TypePtr::BotPTR,ta->ary(),ta->klass(),false,offset);
duke@0 1222 }
duke@0 1223 }
duke@0 1224
duke@0 1225 // Oop pointers need some flattening
duke@0 1226 const TypeInstPtr *to = tj->isa_instptr();
duke@0 1227 if( to && _AliasLevel >= 2 && to != TypeOopPtr::BOTTOM ) {
never@2223 1228 ciInstanceKlass *k = to->klass()->as_instance_klass();
duke@0 1229 if( ptr == TypePtr::Constant ) {
never@2223 1230 if (to->klass() != ciEnv::current()->Class_klass() ||
never@2223 1231 offset < k->size_helper() * wordSize) {
never@2223 1232 // No constant oop pointers (such as Strings); they alias with
never@2223 1233 // unknown strings.
never@2223 1234 assert(!is_known_inst, "not scalarizable allocation");
never@2223 1235 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
never@2223 1236 }
kvn@247 1237 } else if( is_known_inst ) {
kvn@163 1238 tj = to; // Keep NotNull and klass_is_exact for instance type
duke@0 1239 } else if( ptr == TypePtr::NotNull || to->klass_is_exact() ) {
duke@0 1240 // During the 2nd round of IterGVN, NotNull castings are removed.
duke@0 1241 // Make sure the Bottom and NotNull variants alias the same.
duke@0 1242 // Also, make sure exact and non-exact variants alias the same.
kvn@247 1243 tj = to = TypeInstPtr::make(TypePtr::BotPTR,to->klass(),false,0,offset);
duke@0 1244 }
duke@0 1245 // Canonicalize the holder of this field
coleenp@113 1246 if (offset >= 0 && offset < instanceOopDesc::base_offset_in_bytes()) {
duke@0 1247 // First handle header references such as a LoadKlassNode, even if the
duke@0 1248 // object's klass is unloaded at compile time (4965979).
kvn@247 1249 if (!is_known_inst) { // Do it only for non-instance types
kvn@247 1250 tj = to = TypeInstPtr::make(TypePtr::BotPTR, env()->Object_klass(), false, NULL, offset);
kvn@247 1251 }
duke@0 1252 } else if (offset < 0 || offset >= k->size_helper() * wordSize) {
never@2223 1253 // Static fields are in the space above the normal instance
never@2223 1254 // fields in the java.lang.Class instance.
never@2223 1255 if (to->klass() != ciEnv::current()->Class_klass()) {
never@2223 1256 to = NULL;
never@2223 1257 tj = TypeOopPtr::BOTTOM;
never@2223 1258 offset = tj->offset();
never@2223 1259 }
duke@0 1260 } else {
duke@0 1261 ciInstanceKlass *canonical_holder = k->get_canonical_holder(offset);
duke@0 1262 if (!k->equals(canonical_holder) || tj->offset() != offset) {
kvn@247 1263 if( is_known_inst ) {
kvn@247 1264 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, true, NULL, offset, to->instance_id());
kvn@247 1265 } else {
kvn@247 1266 tj = to = TypeInstPtr::make(to->ptr(), canonical_holder, false, NULL, offset);
kvn@247 1267 }
duke@0 1268 }
duke@0 1269 }
duke@0 1270 }
duke@0 1271
duke@0 1272 // Klass pointers to object array klasses need some flattening
duke@0 1273 const TypeKlassPtr *tk = tj->isa_klassptr();
duke@0 1274 if( tk ) {
duke@0 1275 // If we are referencing a field within a Klass, we need
duke@0 1276 // to assume the worst case of an Object. Both exact and
duke@0 1277 // inexact types must flatten to the same alias class.
duke@0 1278 // Since the flattened result for a klass is defined to be
duke@0 1279 // precisely java.lang.Object, use a constant ptr.
duke@0 1280 if ( offset == Type::OffsetBot || (offset >= 0 && (size_t)offset < sizeof(Klass)) ) {
duke@0 1281
duke@0 1282 tj = tk = TypeKlassPtr::make(TypePtr::Constant,
duke@0 1283 TypeKlassPtr::OBJECT->klass(),
duke@0 1284 offset);
duke@0 1285 }
duke@0 1286
duke@0 1287 ciKlass* klass = tk->klass();
duke@0 1288 if( klass->is_obj_array_klass() ) {
duke@0 1289 ciKlass* k = TypeAryPtr::OOPS->klass();
duke@0 1290 if( !k || !k->is_loaded() ) // Only fails for some -Xcomp runs
duke@0 1291 k = TypeInstPtr::BOTTOM->klass();
duke@0 1292 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, k, offset );
duke@0 1293 }
duke@0 1294
duke@0 1295 // Check for precise loads from the primary supertype array and force them
duke@0 1296 // to the supertype cache alias index. Check for generic array loads from
duke@0 1297 // the primary supertype array and also force them to the supertype cache
duke@0 1298 // alias index. Since the same load can reach both, we need to merge
duke@0 1299 // these 2 disparate memories into the same alias class. Since the
duke@0 1300 // primary supertype array is read-only, there's no chance of confusion
duke@0 1301 // where we bypass an array load and an array store.
duke@0 1302 uint off2 = offset - Klass::primary_supers_offset_in_bytes();
duke@0 1303 if( offset == Type::OffsetBot ||
duke@0 1304 off2 < Klass::primary_super_limit()*wordSize ) {
duke@0 1305 offset = sizeof(oopDesc) +Klass::secondary_super_cache_offset_in_bytes();
duke@0 1306 tj = tk = TypeKlassPtr::make( TypePtr::NotNull, tk->klass(), offset );
duke@0 1307 }
duke@0 1308 }
duke@0 1309
duke@0 1310 // Flatten all Raw pointers together.
duke@0 1311 if (tj->base() == Type::RawPtr)
duke@0 1312 tj = TypeRawPtr::BOTTOM;
duke@0 1313
duke@0 1314 if (tj->base() == Type::AnyPtr)
duke@0 1315 tj = TypePtr::BOTTOM; // An error, which the caller must check for.
duke@0 1316
duke@0 1317 // Flatten all to bottom for now
duke@0 1318 switch( _AliasLevel ) {
duke@0 1319 case 0:
duke@0 1320 tj = TypePtr::BOTTOM;
duke@0 1321 break;
duke@0 1322 case 1: // Flatten to: oop, static, field or array
duke@0 1323 switch (tj->base()) {
duke@0 1324 //case Type::AryPtr: tj = TypeAryPtr::RANGE; break;
duke@0 1325 case Type::RawPtr: tj = TypeRawPtr::BOTTOM; break;
duke@0 1326 case Type::AryPtr: // do not distinguish arrays at all
duke@0 1327 case Type::InstPtr: tj = TypeInstPtr::BOTTOM; break;
duke@0 1328 case Type::KlassPtr: tj = TypeKlassPtr::OBJECT; break;
duke@0 1329 case Type::AnyPtr: tj = TypePtr::BOTTOM; break; // caller checks it
duke@0 1330 default: ShouldNotReachHere();
duke@0 1331 }
duke@0 1332 break;
twisti@605 1333 case 2: // No collapsing at level 2; keep all splits
twisti@605 1334 case 3: // No collapsing at level 3; keep all splits
duke@0 1335 break;
duke@0 1336 default:
duke@0 1337 Unimplemented();
duke@0 1338 }
duke@0 1339
duke@0 1340 offset = tj->offset();
duke@0 1341 assert( offset != Type::OffsetTop, "Offset has fallen from constant" );
duke@0 1342
duke@0 1343 assert( (offset != Type::OffsetBot && tj->base() != Type::AryPtr) ||
duke@0 1344 (offset == Type::OffsetBot && tj->base() == Type::AryPtr) ||
duke@0 1345 (offset == Type::OffsetBot && tj == TypeOopPtr::BOTTOM) ||
duke@0 1346 (offset == Type::OffsetBot && tj == TypePtr::BOTTOM) ||
duke@0 1347 (offset == oopDesc::mark_offset_in_bytes() && tj->base() == Type::AryPtr) ||
duke@0 1348 (offset == oopDesc::klass_offset_in_bytes() && tj->base() == Type::AryPtr) ||
duke@0 1349 (offset == arrayOopDesc::length_offset_in_bytes() && tj->base() == Type::AryPtr) ,
duke@0 1350 "For oops, klasses, raw offset must be constant; for arrays the offset is never known" );
duke@0 1351 assert( tj->ptr() != TypePtr::TopPTR &&
duke@0 1352 tj->ptr() != TypePtr::AnyNull &&
duke@0 1353 tj->ptr() != TypePtr::Null, "No imprecise addresses" );
duke@0 1354 // assert( tj->ptr() != TypePtr::Constant ||
duke@0 1355 // tj->base() == Type::RawPtr ||
duke@0 1356 // tj->base() == Type::KlassPtr, "No constant oop addresses" );
duke@0 1357
duke@0 1358 return tj;
duke@0 1359 }
duke@0 1360
duke@0 1361 void Compile::AliasType::Init(int i, const TypePtr* at) {
duke@0 1362 _index = i;
duke@0 1363 _adr_type = at;
duke@0 1364 _field = NULL;
duke@0 1365 _is_rewritable = true; // default
duke@0 1366 const TypeOopPtr *atoop = (at != NULL) ? at->isa_oopptr() : NULL;
kvn@223 1367 if (atoop != NULL && atoop->is_known_instance()) {
kvn@223 1368 const TypeOopPtr *gt = atoop->cast_to_instance_id(TypeOopPtr::InstanceBot);
duke@0 1369 _general_index = Compile::current()->get_alias_index(gt);
duke@0 1370 } else {
duke@0 1371 _general_index = 0;
duke@0 1372 }
duke@0 1373 }
duke@0 1374
duke@0 1375 //---------------------------------print_on------------------------------------
duke@0 1376 #ifndef PRODUCT
duke@0 1377 void Compile::AliasType::print_on(outputStream* st) {
duke@0 1378 if (index() < 10)
duke@0 1379 st->print("@ <%d> ", index());
duke@0 1380 else st->print("@ <%d>", index());
duke@0 1381 st->print(is_rewritable() ? " " : " RO");
duke@0 1382 int offset = adr_type()->offset();
duke@0 1383 if (offset == Type::OffsetBot)
duke@0 1384 st->print(" +any");
duke@0 1385 else st->print(" +%-3d", offset);
duke@0 1386 st->print(" in ");
duke@0 1387 adr_type()->dump_on(st);
duke@0 1388 const TypeOopPtr* tjp = adr_type()->isa_oopptr();
duke@0 1389 if (field() != NULL && tjp) {
duke@0 1390 if (tjp->klass() != field()->holder() ||
duke@0 1391 tjp->offset() != field()->offset_in_bytes()) {
duke@0 1392 st->print(" != ");
duke@0 1393 field()->print();
duke@0 1394 st->print(" ***");
duke@0 1395 }
duke@0 1396 }
duke@0 1397 }
duke@0 1398
duke@0 1399 void print_alias_types() {
duke@0 1400 Compile* C = Compile::current();
duke@0 1401 tty->print_cr("--- Alias types, AliasIdxBot .. %d", C->num_alias_types()-1);
duke@0 1402 for (int idx = Compile::AliasIdxBot; idx < C->num_alias_types(); idx++) {
duke@0 1403 C->alias_type(idx)->print_on(tty);
duke@0 1404 tty->cr();
duke@0 1405 }
duke@0 1406 }
duke@0 1407 #endif
duke@0 1408
duke@0 1409
duke@0 1410 //----------------------------probe_alias_cache--------------------------------
duke@0 1411 Compile::AliasCacheEntry* Compile::probe_alias_cache(const TypePtr* adr_type) {
duke@0 1412 intptr_t key = (intptr_t) adr_type;
duke@0 1413 key ^= key >> logAliasCacheSize;
duke@0 1414 return &_alias_cache[key & right_n_bits(logAliasCacheSize)];
duke@0 1415 }
duke@0 1416
duke@0 1417
duke@0 1418 //-----------------------------grow_alias_types--------------------------------
duke@0 1419 void Compile::grow_alias_types() {
duke@0 1420 const int old_ats = _max_alias_types; // how many before?
duke@0 1421 const int new_ats = old_ats; // how many more?
duke@0 1422 const int grow_ats = old_ats+new_ats; // how many now?
duke@0 1423 _max_alias_types = grow_ats;
duke@0 1424 _alias_types = REALLOC_ARENA_ARRAY(comp_arena(), AliasType*, _alias_types, old_ats, grow_ats);
duke@0 1425 AliasType* ats = NEW_ARENA_ARRAY(comp_arena(), AliasType, new_ats);
duke@0 1426 Copy::zero_to_bytes(ats, sizeof(AliasType)*new_ats);
duke@0 1427 for (int i = 0; i < new_ats; i++) _alias_types[old_ats+i] = &ats[i];
duke@0 1428 }
duke@0 1429
duke@0 1430
duke@0 1431 //--------------------------------find_alias_type------------------------------
never@2223 1432 Compile::AliasType* Compile::find_alias_type(const TypePtr* adr_type, bool no_create, ciField* original_field) {
duke@0 1433 if (_AliasLevel == 0)
duke@0 1434 return alias_type(AliasIdxBot);
duke@0 1435
duke@0 1436 AliasCacheEntry* ace = probe_alias_cache(adr_type);
duke@0 1437 if (ace->_adr_type == adr_type) {
duke@0 1438 return alias_type(ace->_index);
duke@0 1439 }
duke@0 1440
duke@0 1441 // Handle special cases.
duke@0 1442 if (adr_type == NULL) return alias_type(AliasIdxTop);
duke@0 1443 if (adr_type == TypePtr::BOTTOM) return alias_type(AliasIdxBot);
duke@0 1444
duke@0 1445 // Do it the slow way.
duke@0 1446 const TypePtr* flat = flatten_alias_type(adr_type);
duke@0 1447
duke@0 1448 #ifdef ASSERT
duke@0 1449 assert(flat == flatten_alias_type(flat), "idempotent");
duke@0 1450 assert(flat != TypePtr::BOTTOM, "cannot alias-analyze an untyped ptr");
duke@0 1451 if (flat->isa_oopptr() && !flat->isa_klassptr()) {
duke@0 1452 const TypeOopPtr* foop = flat->is_oopptr();
kvn@247 1453 // Scalarizable allocations have exact klass always.
kvn@247 1454 bool exact = !foop->klass_is_exact() || foop->is_known_instance();
kvn@247 1455 const TypePtr* xoop = foop->cast_to_exactness(exact)->is_ptr();
duke@0 1456 assert(foop == flatten_alias_type(xoop), "exactness must not affect alias type");
duke@0 1457 }
duke@0 1458 assert(flat == flatten_alias_type(flat), "exact bit doesn't matter");
duke@0 1459 #endif
duke@0 1460
duke@0 1461 int idx = AliasIdxTop;
duke@0 1462 for (int i = 0; i < num_alias_types(); i++) {
duke@0 1463 if (alias_type(i)->adr_type() == flat) {
duke@0 1464 idx = i;
duke@0 1465 break;
duke@0 1466 }
duke@0 1467 }
duke@0 1468
duke@0 1469 if (idx == AliasIdxTop) {
duke@0 1470 if (no_create) return NULL;
duke@0 1471 // Grow the array if necessary.
duke@0 1472 if (_num_alias_types == _max_alias_types) grow_alias_types();
duke@0 1473 // Add a new alias type.
duke@0 1474 idx = _num_alias_types++;
duke@0 1475 _alias_types[idx]->Init(idx, flat);
duke@0 1476 if (flat == TypeInstPtr::KLASS) alias_type(idx)->set_rewritable(false);
duke@0 1477 if (flat == TypeAryPtr::RANGE) alias_type(idx)->set_rewritable(false);
duke@0 1478 if (flat->isa_instptr()) {
duke@0 1479 if (flat->offset() == java_lang_Class::klass_offset_in_bytes()
duke@0 1480 && flat->is_instptr()->klass() == env()->Class_klass())
duke@0 1481 alias_type(idx)->set_rewritable(false);
duke@0 1482 }
duke@0 1483 if (flat->isa_klassptr()) {
duke@0 1484 if (flat->offset() == Klass::super_check_offset_offset_in_bytes() + (int)sizeof(oopDesc))
duke@0 1485 alias_type(idx)->set_rewritable(false);
duke@0 1486 if (flat->offset() == Klass::modifier_flags_offset_in_bytes() + (int)sizeof(oopDesc))
duke@0 1487 alias_type(idx)->set_rewritable(false);
duke@0 1488 if (flat->offset() == Klass::access_flags_offset_in_bytes() + (int)sizeof(oopDesc))
duke@0 1489 alias_type(idx)->set_rewritable(false);
duke@0 1490 if (flat->offset() == Klass::java_mirror_offset_in_bytes() + (int)sizeof(oopDesc))
duke@0 1491 alias_type(idx)->set_rewritable(false);
duke@0 1492 }
duke@0 1493 // %%% (We would like to finalize JavaThread::threadObj_offset(),
duke@0 1494 // but the base pointer type is not distinctive enough to identify
duke@0 1495 // references into JavaThread.)
duke@0 1496
never@2223 1497 // Check for final fields.
duke@0 1498 const TypeInstPtr* tinst = flat->isa_instptr();
coleenp@113 1499 if (tinst && tinst->offset() >= instanceOopDesc::base_offset_in_bytes()) {
never@2223 1500 ciField* field;
never@2223 1501 if (tinst->const_oop() != NULL &&
never@2223 1502 tinst->klass() == ciEnv::current()->Class_klass() &&
never@2223 1503 tinst->offset() >= (tinst->klass()->as_instance_klass()->size_helper() * wordSize)) {
never@2223 1504 // static field
never@2223 1505 ciInstanceKlass* k = tinst->const_oop()->as_instance()->java_lang_Class_klass()->as_instance_klass();
never@2223 1506 field = k->get_field_by_offset(tinst->offset(), true);
never@2223 1507 } else {
never@2223 1508 ciInstanceKlass *k = tinst->klass()->as_instance_klass();
never@2223 1509 field = k->get_field_by_offset(tinst->offset(), false);
never@2223 1510 }
never@2223 1511 assert(field == NULL ||
never@2223 1512 original_field == NULL ||
never@2223 1513 (field->holder() == original_field->holder() &&
never@2223 1514 field->offset() == original_field->offset() &&
never@2223 1515 field->is_static() == original_field->is_static()), "wrong field?");
duke@0 1516 // Set field() and is_rewritable() attributes.
duke@0 1517 if (field != NULL) alias_type(idx)->set_field(field);
duke@0 1518 }
duke@0 1519 }
duke@0 1520
duke@0 1521 // Fill the cache for next time.
duke@0 1522 ace->_adr_type = adr_type;
duke@0 1523 ace->_index = idx;
duke@0 1524 assert(alias_type(adr_type) == alias_type(idx), "type must be installed");
duke@0 1525
duke@0 1526 // Might as well try to fill the cache for the flattened version, too.
duke@0 1527 AliasCacheEntry* face = probe_alias_cache(flat);
duke@0 1528 if (face->_adr_type == NULL) {
duke@0 1529 face->_adr_type = flat;
duke@0 1530 face->_index = idx;
duke@0 1531 assert(alias_type(flat) == alias_type(idx), "flat type must work too");
duke@0 1532 }
duke@0 1533
duke@0 1534 return alias_type(idx);
duke@0 1535 }
duke@0 1536
duke@0 1537
duke@0 1538 Compile::AliasType* Compile::alias_type(ciField* field) {
duke@0 1539 const TypeOopPtr* t;
duke@0 1540 if (field->is_static())
never@2223 1541 t = TypeInstPtr::make(field->holder()->java_mirror());
duke@0 1542 else
duke@0 1543 t = TypeOopPtr::make_from_klass_raw(field->holder());
never@2223 1544 AliasType* atp = alias_type(t->add_offset(field->offset_in_bytes()), field);
duke@0 1545 assert(field->is_final() == !atp->is_rewritable(), "must get the rewritable bits correct");
duke@0 1546 return atp;
duke@0 1547 }
duke@0 1548
duke@0 1549
duke@0 1550 //------------------------------have_alias_type--------------------------------
duke@0 1551 bool Compile::have_alias_type(const TypePtr* adr_type) {
duke@0 1552 AliasCacheEntry* ace = probe_alias_cache(adr_type);
duke@0 1553 if (ace->_adr_type == adr_type) {
duke@0 1554 return true;
duke@0 1555 }
duke@0 1556
duke@0 1557 // Handle special cases.
duke@0 1558 if (adr_type == NULL) return true;
duke@0 1559 if (adr_type == TypePtr::BOTTOM) return true;
duke@0 1560
never@2223 1561 return find_alias_type(adr_type, true, NULL) != NULL;
duke@0 1562 }
duke@0 1563
duke@0 1564 //-----------------------------must_alias--------------------------------------
duke@0 1565 // True if all values of the given address type are in the given alias category.
duke@0 1566 bool Compile::must_alias(const TypePtr* adr_type, int alias_idx) {
duke@0 1567 if (alias_idx == AliasIdxBot) return true; // the universal category
duke@0 1568 if (adr_type == NULL) return true; // NULL serves as TypePtr::TOP
duke@0 1569 if (alias_idx == AliasIdxTop) return false; // the empty category
duke@0 1570 if (adr_type->base() == Type::AnyPtr) return false; // TypePtr::BOTTOM or its twins
duke@0 1571
duke@0 1572 // the only remaining possible overlap is identity
duke@0 1573 int adr_idx = get_alias_index(adr_type);
duke@0 1574 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
duke@0 1575 assert(adr_idx == alias_idx ||
duke@0 1576 (alias_type(alias_idx)->adr_type() != TypeOopPtr::BOTTOM
duke@0 1577 && adr_type != TypeOopPtr::BOTTOM),
duke@0 1578 "should not be testing for overlap with an unsafe pointer");
duke@0 1579 return adr_idx == alias_idx;
duke@0 1580 }
duke@0 1581
duke@0 1582 //------------------------------can_alias--------------------------------------
duke@0 1583 // True if any values of the given address type are in the given alias category.
duke@0 1584 bool Compile::can_alias(const TypePtr* adr_type, int alias_idx) {
duke@0 1585 if (alias_idx == AliasIdxTop) return false; // the empty category
duke@0 1586 if (adr_type == NULL) return false; // NULL serves as TypePtr::TOP
duke@0 1587 if (alias_idx == AliasIdxBot) return true; // the universal category
duke@0 1588 if (adr_type->base() == Type::AnyPtr) return true; // TypePtr::BOTTOM or its twins
duke@0 1589
duke@0 1590 // the only remaining possible overlap is identity
duke@0 1591 int adr_idx = get_alias_index(adr_type);
duke@0 1592 assert(adr_idx != AliasIdxBot && adr_idx != AliasIdxTop, "");
duke@0 1593 return adr_idx == alias_idx;
duke@0 1594 }
duke@0 1595
duke@0 1596
duke@0 1597
duke@0 1598 //---------------------------pop_warm_call-------------------------------------
duke@0 1599 WarmCallInfo* Compile::pop_warm_call() {
duke@0 1600 WarmCallInfo* wci = _warm_calls;
duke@0 1601 if (wci != NULL) _warm_calls = wci->remove_from(wci);
duke@0 1602 return wci;
duke@0 1603 }
duke@0 1604
duke@0 1605 //----------------------------Inline_Warm--------------------------------------
duke@0 1606 int Compile::Inline_Warm() {
duke@0 1607 // If there is room, try to inline some more warm call sites.
duke@0 1608 // %%% Do a graph index compaction pass when we think we're out of space?
duke@0 1609 if (!InlineWarmCalls) return 0;
duke@0 1610
duke@0 1611 int calls_made_hot = 0;
duke@0 1612 int room_to_grow = NodeCountInliningCutoff - unique();
duke@0 1613 int amount_to_grow = MIN2(room_to_grow, (int)NodeCountInliningStep);
duke@0 1614 int amount_grown = 0;
duke@0 1615 WarmCallInfo* call;
duke@0 1616 while (amount_to_grow > 0 && (call = pop_warm_call()) != NULL) {
duke@0 1617 int est_size = (int)call->size();
duke@0 1618 if (est_size > (room_to_grow - amount_grown)) {
duke@0 1619 // This one won't fit anyway. Get rid of it.
duke@0 1620 call->make_cold();
duke@0 1621 continue;
duke@0 1622 }
duke@0 1623 call->make_hot();
duke@0 1624 calls_made_hot++;
duke@0 1625 amount_grown += est_size;
duke@0 1626 amount_to_grow -= est_size;
duke@0 1627 }
duke@0 1628
duke@0 1629 if (calls_made_hot > 0) set_major_progress();
duke@0 1630 return calls_made_hot;
duke@0 1631 }
duke@0 1632
duke@0 1633
duke@0 1634 //----------------------------Finish_Warm--------------------------------------
duke@0 1635 void Compile::Finish_Warm() {
duke@0 1636 if (!InlineWarmCalls) return;
duke@0 1637 if (failing()) return;
duke@0 1638 if (warm_calls() == NULL) return;
duke@0 1639
duke@0 1640 // Clean up loose ends, if we are out of space for inlining.
duke@0 1641 WarmCallInfo* call;
duke@0 1642 while ((call = pop_warm_call()) != NULL) {
duke@0 1643 call->make_cold();
duke@0 1644 }
duke@0 1645 }
duke@0 1646
cfang@1172 1647 //---------------------cleanup_loop_predicates-----------------------
cfang@1172 1648 // Remove the opaque nodes that protect the predicates so that all unused
cfang@1172 1649 // checks and uncommon_traps will be eliminated from the ideal graph
cfang@1172 1650 void Compile::cleanup_loop_predicates(PhaseIterGVN &igvn) {
cfang@1172 1651 if (predicate_count()==0) return;
cfang@1172 1652 for (int i = predicate_count(); i > 0; i--) {
cfang@1172 1653 Node * n = predicate_opaque1_node(i-1);
cfang@1172 1654 assert(n->Opcode() == Op_Opaque1, "must be");
cfang@1172 1655 igvn.replace_node(n, n->in(1));
cfang@1172 1656 }
cfang@1172 1657 assert(predicate_count()==0, "should be clean!");
cfang@1172 1658 }
duke@0 1659
duke@0 1660 //------------------------------Optimize---------------------------------------
duke@0 1661 // Given a graph, optimize it.
duke@0 1662 void Compile::Optimize() {
duke@0 1663 TracePhase t1("optimizer", &_t_optimizer, true);
duke@0 1664
duke@0 1665 #ifndef PRODUCT
duke@0 1666 if (env()->break_at_compile()) {
duke@0 1667 BREAKPOINT;
duke@0 1668 }
duke@0 1669
duke@0 1670 #endif
duke@0 1671
duke@0 1672 ResourceMark rm;
duke@0 1673 int loop_opts_cnt;
duke@0 1674
duke@0 1675 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1676
never@222 1677 print_method("After Parsing");
duke@0 1678
duke@0 1679 {
duke@0 1680 // Iterative Global Value Numbering, including ideal transforms
duke@0 1681 // Initialize IterGVN with types and values from parse-time GVN
duke@0 1682 PhaseIterGVN igvn(initial_gvn());
duke@0 1683 {
duke@0 1684 NOT_PRODUCT( TracePhase t2("iterGVN", &_t_iterGVN, TimeCompiler); )
duke@0 1685 igvn.optimize();
duke@0 1686 }
duke@0 1687
duke@0 1688 print_method("Iter GVN 1", 2);
duke@0 1689
duke@0 1690 if (failing()) return;
duke@0 1691
kvn@1554 1692 // Perform escape analysis
kvn@1554 1693 if (_do_escape_analysis && ConnectionGraph::has_candidates(this)) {
kvn@1554 1694 TracePhase t2("escapeAnalysis", &_t_escapeAnalysis, true);
kvn@1554 1695 ConnectionGraph::do_analysis(this, &igvn);
kvn@1554 1696
kvn@1554 1697 if (failing()) return;
kvn@1554 1698
kvn@1554 1699 igvn.optimize();
kvn@1554 1700 print_method("Iter GVN 3", 2);
kvn@1554 1701
kvn@1554 1702 if (failing()) return;
kvn@1554 1703
kvn@1554 1704 }
kvn@1554 1705
duke@0 1706 // Loop transforms on the ideal graph. Range Check Elimination,
duke@0 1707 // peeling, unrolling, etc.
duke@0 1708
duke@0 1709 // Set loop opts counter
duke@0 1710 loop_opts_cnt = num_loop_opts();
duke@0 1711 if((loop_opts_cnt > 0) && (has_loops() || has_split_ifs())) {
duke@0 1712 {
duke@0 1713 TracePhase t2("idealLoop", &_t_idealLoop, true);
kvn@2292 1714 PhaseIdealLoop ideal_loop( igvn, true );
duke@0 1715 loop_opts_cnt--;
duke@0 1716 if (major_progress()) print_method("PhaseIdealLoop 1", 2);
duke@0 1717 if (failing()) return;
duke@0 1718 }
duke@0 1719 // Loop opts pass if partial peeling occurred in previous pass
duke@0 1720 if(PartialPeelLoop && major_progress() && (loop_opts_cnt > 0)) {
duke@0 1721 TracePhase t3("idealLoop", &_t_idealLoop, true);
kvn@2292 1722 PhaseIdealLoop ideal_loop( igvn, false );
duke@0 1723 loop_opts_cnt--;
duke@0 1724 if (major_progress()) print_method("PhaseIdealLoop 2", 2);
duke@0 1725 if (failing()) return;
duke@0 1726 }
duke@0 1727 // Loop opts pass for loop-unrolling before CCP
duke@0 1728 if(major_progress() && (loop_opts_cnt > 0)) {
duke@0 1729 TracePhase t4("idealLoop", &_t_idealLoop, true);
kvn@2292 1730 PhaseIdealLoop ideal_loop( igvn, false );
duke@0 1731 loop_opts_cnt--;
duke@0 1732 if (major_progress()) print_method("PhaseIdealLoop 3", 2);
duke@0 1733 }
never@921 1734 if (!failing()) {
never@921 1735 // Verify that last round of loop opts produced a valid graph
never@921 1736 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
never@921 1737 PhaseIdealLoop::verify(igvn);
never@921 1738 }
duke@0 1739 }
duke@0 1740 if (failing()) return;
duke@0 1741
duke@0 1742 // Conditional Constant Propagation;
duke@0 1743 PhaseCCP ccp( &igvn );
duke@0 1744 assert( true, "Break here to ccp.dump_nodes_and_types(_root,999,1)");
duke@0 1745 {
duke@0 1746 TracePhase t2("ccp", &_t_ccp, true);
duke@0 1747 ccp.do_transform();
duke@0 1748 }
duke@0 1749 print_method("PhaseCPP 1", 2);
duke@0 1750
duke@0 1751 assert( true, "Break here to ccp.dump_old2new_map()");
duke@0 1752
duke@0 1753 // Iterative Global Value Numbering, including ideal transforms
duke@0 1754 {
duke@0 1755 NOT_PRODUCT( TracePhase t2("iterGVN2", &_t_iterGVN2, TimeCompiler); )
duke@0 1756 igvn = ccp;
duke@0 1757 igvn.optimize();
duke@0 1758 }
duke@0 1759
duke@0 1760 print_method("Iter GVN 2", 2);
duke@0 1761
duke@0 1762 if (failing()) return;
duke@0 1763
duke@0 1764 // Loop transforms on the ideal graph. Range Check Elimination,
duke@0 1765 // peeling, unrolling, etc.
duke@0 1766 if(loop_opts_cnt > 0) {
duke@0 1767 debug_only( int cnt = 0; );
duke@0 1768 while(major_progress() && (loop_opts_cnt > 0)) {
duke@0 1769 TracePhase t2("idealLoop", &_t_idealLoop, true);
duke@0 1770 assert( cnt++ < 40, "infinite cycle in loop optimization" );
kvn@2292 1771 PhaseIdealLoop ideal_loop( igvn, true);
duke@0 1772 loop_opts_cnt--;
duke@0 1773 if (major_progress()) print_method("PhaseIdealLoop iterations", 2);
duke@0 1774 if (failing()) return;
duke@0 1775 }
duke@0 1776 }
never@921 1777
never@921 1778 {
never@921 1779 // Verify that all previous optimizations produced a valid graph
never@921 1780 // at least to this point, even if no loop optimizations were done.
never@921 1781 NOT_PRODUCT( TracePhase t2("idealLoopVerify", &_t_idealLoopVerify, TimeCompiler); )
never@921 1782 PhaseIdealLoop::verify(igvn);
never@921 1783 }
never@921 1784
duke@0 1785 {
duke@0 1786 NOT_PRODUCT( TracePhase t2("macroExpand", &_t_macroExpand, TimeCompiler); )
duke@0 1787 PhaseMacroExpand mex(igvn);
duke@0 1788 if (mex.expand_macro_nodes()) {
duke@0 1789 assert(failing(), "must bail out w/ explicit message");
duke@0 1790 return;
duke@0 1791 }
duke@0 1792 }
duke@0 1793
duke@0 1794 } // (End scope of igvn; run destructor if necessary for asserts.)
duke@0 1795
duke@0 1796 // A method with only infinite loops has no edges entering loops from root
duke@0 1797 {
duke@0 1798 NOT_PRODUCT( TracePhase t2("graphReshape", &_t_graphReshaping, TimeCompiler); )
duke@0 1799 if (final_graph_reshaping()) {
duke@0 1800 assert(failing(), "must bail out w/ explicit message");
duke@0 1801 return;
duke@0 1802 }
duke@0 1803 }
duke@0 1804
duke@0 1805 print_method("Optimize finished", 2);
duke@0 1806 }
duke@0 1807
duke@0 1808
duke@0 1809 //------------------------------Code_Gen---------------------------------------
duke@0 1810 // Given a graph, generate code for it
duke@0 1811 void Compile::Code_Gen() {
duke@0 1812 if (failing()) return;
duke@0 1813
duke@0 1814 // Perform instruction selection. You might think we could reclaim Matcher
duke@0 1815 // memory PDQ, but actually the Matcher is used in generating spill code.
duke@0 1816 // Internals of the Matcher (including some VectorSets) must remain live
duke@0 1817 // for awhile - thus I cannot reclaim Matcher memory lest a VectorSet usage
duke@0 1818 // set a bit in reclaimed memory.
duke@0 1819
duke@0 1820 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
duke@0 1821 // nodes. Mapping is only valid at the root of each matched subtree.
duke@0 1822 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1823
duke@0 1824 Node_List proj_list;
duke@0 1825 Matcher m(proj_list);
duke@0 1826 _matcher = &m;
duke@0 1827 {
duke@0 1828 TracePhase t2("matcher", &_t_matcher, true);
duke@0 1829 m.match();
duke@0 1830 }
duke@0 1831 // In debug mode can dump m._nodes.dump() for mapping of ideal to machine
duke@0 1832 // nodes. Mapping is only valid at the root of each matched subtree.
duke@0 1833 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1834
duke@0 1835 // If you have too many nodes, or if matching has failed, bail out
duke@0 1836 check_node_count(0, "out of nodes matching instructions");
duke@0 1837 if (failing()) return;
duke@0 1838
duke@0 1839 // Build a proper-looking CFG
duke@0 1840 PhaseCFG cfg(node_arena(), root(), m);
duke@0 1841 _cfg = &cfg;
duke@0 1842 {
duke@0 1843 NOT_PRODUCT( TracePhase t2("scheduler", &_t_scheduler, TimeCompiler); )
duke@0 1844 cfg.Dominators();
duke@0 1845 if (failing()) return;
duke@0 1846
duke@0 1847 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1848
duke@0 1849 cfg.Estimate_Block_Frequency();
duke@0 1850 cfg.GlobalCodeMotion(m,unique(),proj_list);
duke@0 1851
duke@0 1852 print_method("Global code motion", 2);
duke@0 1853
duke@0 1854 if (failing()) return;
duke@0 1855 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1856
duke@0 1857 debug_only( cfg.verify(); )
duke@0 1858 }
duke@0 1859 NOT_PRODUCT( verify_graph_edges(); )
duke@0 1860
duke@0 1861 PhaseChaitin regalloc(unique(),cfg,m);
duke@0 1862 _regalloc = &regalloc;
duke@0 1863 {
duke@0 1864 TracePhase t2("regalloc", &_t_registerAllocation, true);
duke@0 1865 // Perform any platform dependent preallocation actions. This is used,
duke@0 1866 // for example, to avoid taking an implicit null pointer exception
duke@0 1867 // using the frame pointer on win95.
duke@0 1868 _regalloc->pd_preallocate_hook();
duke@0 1869
duke@0 1870 // Perform register allocation. After Chaitin, use-def chains are
duke@0 1871 // no longer accurate (at spill code) and so must be ignored.
duke@0 1872 // Node->LRG->reg mappings are still accurate.
duke@0 1873 _regalloc->Register_Allocate();
duke@0 1874
duke@0 1875 // Bail out if the allocator builds too many nodes
duke@0 1876 if (failing()) return;
duke@0 1877 }
duke@0 1878
duke@0 1879 // Prior to register allocation we kept empty basic blocks in case the
duke@0 1880 // the allocator needed a place to spill. After register allocation we
duke@0 1881 // are not adding any new instructions. If any basic block is empty, we
duke@0 1882 // can now safely remove it.
duke@0 1883 {
rasbold@418 1884 NOT_PRODUCT( TracePhase t2("blockOrdering", &_t_blockOrdering, TimeCompiler); )
rasbold@418 1885 cfg.remove_empty();
rasbold@418 1886 if (do_freq_based_layout()) {
rasbold@418 1887 PhaseBlockLayout layout(cfg);
rasbold@418 1888 } else {
rasbold@418 1889 cfg.set_loop_alignment();
rasbold@418 1890 }
rasbold@418 1891 cfg.fixup_flow();
duke@0 1892 }
duke@0 1893
duke@0 1894 // Perform any platform dependent postallocation verifications.
duke@0 1895 debug_only( _regalloc->pd_postallocate_verify_hook(); )
duke@0 1896
duke@0 1897 // Apply peephole optimizations
duke@0 1898 if( OptoPeephole ) {
duke@0 1899 NOT_PRODUCT( TracePhase t2("peephole", &_t_peephole, TimeCompiler); )
duke@0 1900 PhasePeephole peep( _regalloc, cfg);
duke@0 1901 peep.do_transform();
duke@0 1902 }
duke@0 1903
duke@0 1904 // Convert Nodes to instruction bits in a buffer
duke@0 1905 {
duke@0 1906 // %%%% workspace merge brought two timers together for one job
duke@0 1907 TracePhase t2a("output", &_t_output, true);
duke@0 1908 NOT_PRODUCT( TraceTime t2b(NULL, &_t_codeGeneration, TimeCompiler, false); )
duke@0 1909 Output();
duke@0 1910 }
duke@0 1911
never@222 1912 print_method("Final Code");
duke@0 1913
duke@0 1914 // He's dead, Jim.
duke@0 1915 _cfg = (PhaseCFG*)0xdeadbeef;
duke@0 1916 _regalloc = (PhaseChaitin*)0xdeadbeef;
duke@0 1917 }
duke@0 1918
duke@0 1919
duke@0 1920 //------------------------------dump_asm---------------------------------------
duke@0 1921 // Dump formatted assembly
duke@0 1922 #ifndef PRODUCT
duke@0 1923 void Compile::dump_asm(int *pcs, uint pc_limit) {
duke@0 1924 bool cut_short = false;
duke@0 1925 tty->print_cr("#");
duke@0 1926 tty->print("# "); _tf->dump(); tty->cr();
duke@0 1927 tty->print_cr("#");
duke@0 1928
duke@0 1929 // For all blocks
duke@0 1930 int pc = 0x0; // Program counter
duke@0 1931 char starts_bundle = ' ';
duke@0 1932 _regalloc->dump_frame();
duke@0 1933
duke@0 1934 Node *n = NULL;
duke@0 1935 for( uint i=0; i<_cfg->_num_blocks; i++ ) {
duke@0 1936 if (VMThread::should_terminate()) { cut_short = true; break; }
duke@0 1937 Block *b = _cfg->_blocks[i];
duke@0 1938 if (b->is_connector() && !Verbose) continue;
duke@0 1939 n = b->_nodes[0];
duke@0 1940 if (pcs && n->_idx < pc_limit)
duke@0 1941 tty->print("%3.3x ", pcs[n->_idx]);
duke@0 1942 else
duke@0 1943 tty->print(" ");
duke@0 1944 b->dump_head( &_cfg->_bbs );
duke@0 1945 if (b->is_connector()) {
duke@0 1946 tty->print_cr(" # Empty connector block");
duke@0 1947 } else if (b->num_preds() == 2 && b->pred(1)->is_CatchProj() && b->pred(1)->as_CatchProj()->_con == CatchProjNode::fall_through_index) {
duke@0 1948 tty->print_cr(" # Block is sole successor of call");
duke@0 1949 }
duke@0 1950
duke@0 1951 // For all instructions
duke@0 1952 Node *delay = NULL;
duke@0 1953 for( uint j = 0; j<b->_nodes.size(); j++ ) {
duke@0 1954 if (VMThread::should_terminate()) { cut_short = true; break; }
duke@0 1955 n = b->_nodes[j];
duke@0 1956 if (valid_bundle_info(n)) {
duke@0 1957 Bundle *bundle = node_bundling(n);
duke@0 1958 if (bundle->used_in_unconditional_delay()) {
duke@0 1959 delay = n;
duke@0 1960 continue;
duke@0 1961 }
duke@0 1962 if (bundle->starts_bundle())
duke@0 1963 starts_bundle = '+';
duke@0 1964 }
duke@0 1965
coleenp@113 1966 if (WizardMode) n->dump();
coleenp@113 1967
duke@0 1968 if( !n->is_Region() && // Dont print in the Assembly
duke@0 1969 !n->is_Phi() && // a few noisely useless nodes
duke@0 1970 !n->is_Proj() &&
duke@0 1971 !n->is_MachTemp() &&
kvn@1100 1972 !n->is_SafePointScalarObject() &&
duke@0 1973 !n->is_Catch() && // Would be nice to print exception table targets
duke@0 1974 !n->is_MergeMem() && // Not very interesting
duke@0 1975 !n->is_top() && // Debug info table constants
duke@0 1976 !(n->is_Con() && !n->is_Mach())// Debug info table constants
duke@0 1977 ) {
duke@0 1978 if (pcs && n->_idx < pc_limit)
duke@0 1979 tty->print("%3.3x", pcs[n->_idx]);
duke@0 1980 else
duke@0 1981 tty->print(" ");
duke@0 1982 tty->print(" %c ", starts_bundle);
duke@0 1983 starts_bundle = ' ';
duke@0 1984 tty->print("\t");
duke@0 1985 n->format(_regalloc, tty);
duke@0 1986 tty->cr();
duke@0 1987 }
duke@0 1988
duke@0 1989 // If we have an instruction with a delay slot, and have seen a delay,
duke@0 1990 // then back up and print it
duke@0 1991 if (valid_bundle_info(n) && node_bundling(n)->use_unconditional_delay()) {
duke@0 1992 assert(delay != NULL, "no unconditional delay instruction");
coleenp@113 1993 if (WizardMode) delay->dump();
coleenp@113 1994
duke@0 1995 if (node_bundling(delay)->starts_bundle())
duke@0 1996 starts_bundle = '+';
duke@0 1997 if (pcs && n->_idx < pc_limit)
duke@0 1998 tty->print("%3.3x", pcs[n->_idx]);
duke@0 1999 else
duke@0 2000 tty->print(" ");
duke@0 2001 tty->print(" %c ", starts_bundle);
duke@0 2002 starts_bundle = ' ';
duke@0 2003 tty->print("\t");
duke@0 2004 delay->format(_regalloc, tty);
duke@0 2005 tty->print_cr("");
duke@0 2006 delay = NULL;
duke@0 2007 }
duke@0 2008
duke@0 2009 // Dump the exception table as well
duke@0 2010 if( n->is_Catch() && (Verbose || WizardMode) ) {
duke@0 2011 // Print the exception table for this offset
duke@0 2012 _handler_table.print_subtable_for(pc);
duke@0 2013 }
duke@0 2014 }
duke@0 2015
duke@0 2016 if (pcs && n->_idx < pc_limit)
duke@0 2017 tty->print_cr("%3.3x", pcs[n->_idx]);
duke@0 2018 else
duke@0 2019 tty->print_cr("");
duke@0 2020
duke@0 2021 assert(cut_short || delay == NULL, "no unconditional delay branch");
duke@0 2022
duke@0 2023 } // End of per-block dump
duke@0 2024 tty->print_cr("");
duke@0 2025
duke@0 2026 if (cut_short) tty->print_cr("*** disassembly is cut short ***");
duke@0 2027 }
duke@0 2028 #endif
duke@0 2029
duke@0 2030 //------------------------------Final_Reshape_Counts---------------------------
duke@0 2031 // This class defines counters to help identify when a method
duke@0 2032 // may/must be executed using hardware with only 24-bit precision.
duke@0 2033 struct Final_Reshape_Counts : public StackObj {
duke@0 2034 int _call_count; // count non-inlined 'common' calls
duke@0 2035 int _float_count; // count float ops requiring 24-bit precision
duke@0 2036 int _double_count; // count double ops requiring more precision
duke@0 2037 int _java_call_count; // count non-inlined 'java' calls
kvn@859 2038 int _inner_loop_count; // count loops which need alignment
duke@0 2039 VectorSet _visited; // Visitation flags
duke@0 2040 Node_List _tests; // Set of IfNodes & PCTableNodes
duke@0 2041
duke@0 2042 Final_Reshape_Counts() :
kvn@859 2043 _call_count(0), _float_count(0), _double_count(0),
kvn@859 2044 _java_call_count(0), _inner_loop_count(0),
duke@0 2045 _visited( Thread::current()->resource_area() ) { }
duke@0 2046
duke@0 2047 void inc_call_count () { _call_count ++; }
duke@0 2048 void inc_float_count () { _float_count ++; }
duke@0 2049 void inc_double_count() { _double_count++; }
duke@0 2050 void inc_java_call_count() { _java_call_count++; }
kvn@859 2051 void inc_inner_loop_count() { _inner_loop_count++; }
duke@0 2052
duke@0 2053 int get_call_count () const { return _call_count ; }
duke@0 2054 int get_float_count () const { return _float_count ; }
duke@0 2055 int get_double_count() const { return _double_count; }
duke@0 2056 int get_java_call_count() const { return _java_call_count; }
kvn@859 2057 int get_inner_loop_count() const { return _inner_loop_count; }
duke@0 2058 };
duke@0 2059
duke@0 2060 static bool oop_offset_is_sane(const TypeInstPtr* tp) {
duke@0 2061 ciInstanceKlass *k = tp->klass()->as_instance_klass();
duke@0 2062 // Make sure the offset goes inside the instance layout.
coleenp@113 2063 return k->contains_field_offset(tp->offset());
duke@0 2064 // Note that OffsetBot and OffsetTop are very negative.
duke@0 2065 }
duke@0 2066
never@2345 2067 // Eliminate trivially redundant StoreCMs and accumulate their
never@2345 2068 // precedence edges.
never@2345 2069 static void eliminate_redundant_card_marks(Node* n) {
never@2345 2070 assert(n->Opcode() == Op_StoreCM, "expected StoreCM");
never@2345 2071 if (n->in(MemNode::Address)->outcnt() > 1) {
never@2345 2072 // There are multiple users of the same address so it might be
never@2345 2073 // possible to eliminate some of the StoreCMs
never@2345 2074 Node* mem = n->in(MemNode::Memory);
never@2345 2075 Node* adr = n->in(MemNode::Address);
never@2345 2076 Node* val = n->in(MemNode::ValueIn);
never@2345 2077 Node* prev = n;
never@2345 2078 bool done = false;
never@2345 2079 // Walk the chain of StoreCMs eliminating ones that match. As
never@2345 2080 // long as it's a chain of single users then the optimization is
never@2345 2081 // safe. Eliminating partially redundant StoreCMs would require
never@2345 2082 // cloning copies down the other paths.
never@2345 2083 while (mem->Opcode() == Op_StoreCM && mem->outcnt() == 1 && !done) {
never@2345 2084 if (adr == mem->in(MemNode::Address) &&
never@2345 2085 val == mem->in(MemNode::ValueIn)) {
never@2345 2086 // redundant StoreCM
never@2345 2087 if (mem->req() > MemNode::OopStore) {
never@2345 2088 // Hasn't been processed by this code yet.
never@2345 2089 n->add_prec(mem->in(MemNode::OopStore));
never@2345 2090 } else {
never@2345 2091 // Already converted to precedence edge
never@2345 2092 for (uint i = mem->req(); i < mem->len(); i++) {
never@2345 2093 // Accumulate any precedence edges
never@2345 2094 if (mem->in(i) != NULL) {
never@2345 2095 n->add_prec(mem->in(i));
never@2345 2096 }
never@2345 2097 }
never@2345 2098 // Everything above this point has been processed.
never@2345 2099 done = true;
never@2345 2100 }
never@2345 2101 // Eliminate the previous StoreCM
never@2345 2102 prev->set_req(MemNode::Memory, mem->in(MemNode::Memory));
never@2345 2103 assert(mem->outcnt() == 0, "should be dead");
never@2345 2104 mem->disconnect_inputs(NULL);
never@2345 2105 } else {
never@2345 2106 prev = mem;
never@2345 2107 }
never@2345 2108 mem = prev->in(MemNode::Memory);
never@2345 2109 }
never@2345 2110 }
never@2345 2111 }
never@2345 2112
duke@0 2113 //------------------------------final_graph_reshaping_impl----------------------
duke@0 2114 // Implement items 1-5 from final_graph_reshaping below.
kvn@859 2115 static void final_graph_reshaping_impl( Node *n, Final_Reshape_Counts &frc ) {
duke@0 2116
kvn@168 2117 if ( n->outcnt() == 0 ) return; // dead node
duke@0 2118 uint nop = n->Opcode();
duke@0 2119
duke@0 2120 // Check for 2-input instruction with "last use" on right input.
duke@0 2121 // Swap to left input. Implements item (2).
duke@0 2122 if( n->req() == 3 && // two-input instruction
duke@0 2123 n->in(1)->outcnt() > 1 && // left use is NOT a last use
duke@0 2124 (!n->in(1)->is_Phi() || n->in(1)->in(2) != n) && // it is not data loop
duke@0 2125 n->in(2)->outcnt() == 1 &&// right use IS a last use
duke@0 2126 !n->in(2)->is_Con() ) { // right use is not a constant
duke@0 2127 // Check for commutative opcode
duke@0 2128 switch( nop ) {
duke@0 2129 case Op_AddI: case Op_AddF: case Op_AddD: case Op_AddL:
duke@0 2130 case Op_MaxI: case Op_MinI:
duke@0 2131 case Op_MulI: case Op_MulF: case Op_MulD: case Op_MulL:
duke@0 2132 case Op_AndL: case Op_XorL: case Op_OrL:
duke@0 2133 case Op_AndI: case Op_XorI: case Op_OrI: {
duke@0 2134 // Move "last use" input to left by swapping inputs
duke@0 2135 n->swap_edges(1, 2);
duke@0 2136 break;
duke@0 2137 }
duke@0 2138 default:
duke@0 2139 break;
duke@0 2140 }
duke@0 2141 }
duke@0 2142
kvn@1529 2143 #ifdef ASSERT
kvn@1529 2144 if( n->is_Mem() ) {
kvn@1529 2145 Compile* C = Compile::current();
kvn@1529 2146 int alias_idx = C->get_alias_index(n->as_Mem()->adr_type());
kvn@1529 2147 assert( n->in(0) != NULL || alias_idx != Compile::AliasIdxRaw ||
kvn@1529 2148 // oop will be recorded in oop map if load crosses safepoint
kvn@1529 2149 n->is_Load() && (n->as_Load()->bottom_type()->isa_oopptr() ||
kvn@1529 2150 LoadNode::is_immutable_value(n->in(MemNode::Address))),
kvn@1529 2151 "raw memory operations should have control edge");
kvn@1529 2152 }
kvn@1529 2153 #endif
duke@0 2154 // Count FPU ops and common calls, implements item (3)
duke@0 2155 switch( nop ) {
duke@0 2156 // Count all float operations that may use FPU
duke@0 2157 case Op_AddF:
duke@0 2158 case Op_SubF:
duke@0 2159 case Op_MulF:
duke@0 2160 case Op_DivF:
duke@0 2161 case Op_NegF:
duke@0 2162 case Op_ModF:
duke@0 2163 case Op_ConvI2F:
duke@0 2164 case Op_ConF:
duke@0 2165 case Op_CmpF:
duke@0 2166 case Op_CmpF3:
duke@0 2167 // case Op_ConvL2F: // longs are split into 32-bit halves
kvn@859 2168 frc.inc_float_count();
duke@0 2169 break;
duke@0 2170
duke@0 2171 case Op_ConvF2D:
duke@0 2172 case Op_ConvD2F:
kvn@859 2173 frc.inc_float_count();
kvn@859 2174 frc.inc_double_count();
duke@0 2175 break;
duke@0 2176
duke@0 2177 // Count all double operations that may use FPU
duke@0 2178 case Op_AddD:
duke@0 2179 case Op_SubD:
duke@0 2180 case Op_MulD:
duke@0 2181 case Op_DivD:
duke@0 2182 case Op_NegD:
duke@0 2183 case Op_ModD:
duke@0 2184 case Op_ConvI2D:
duke@0 2185 case Op_ConvD2I:
duke@0 2186 // case Op_ConvL2D: // handled by leaf call
duke@0 2187 // case Op_ConvD2L: // handled by leaf call
duke@0 2188 case Op_ConD:
duke@0 2189 case Op_CmpD:
duke@0 2190 case Op_CmpD3:
kvn@859 2191 frc.inc_double_count();
duke@0 2192 break;
duke@0 2193 case Op_Opaque1: // Remove Opaque Nodes before matching
duke@0 2194 case Op_Opaque2: // Remove Opaque Nodes before matching
kvn@168 2195 n->subsume_by(n->in(1));
duke@0 2196 break;
duke@0 2197 case Op_CallStaticJava:
duke@0 2198 case Op_CallJava:
duke@0 2199 case Op_CallDynamicJava:
kvn@859 2200 frc.inc_java_call_count(); // Count java call site;
duke@0 2201 case Op_CallRuntime:
duke@0 2202 case Op_CallLeaf:
duke@0 2203 case Op_CallLeafNoFP: {
duke@0 2204 assert( n->is_Call(), "" );
duke@0 2205 CallNode *call = n->as_Call();
duke@0 2206 // Count call sites where the FP mode bit would have to be flipped.
duke@0 2207 // Do not count uncommon runtime calls:
duke@0 2208 // uncommon_trap, _complete_monitor_locking, _complete_monitor_unlocking,
duke@0 2209 // _new_Java, _new_typeArray, _new_objArray, _rethrow_Java, ...
duke@0 2210 if( !call->is_CallStaticJava() || !call->as_CallStaticJava()->_name ) {
kvn@859 2211 frc.inc_call_count(); // Count the call site
duke@0 2212 } else { // See if uncommon argument is shared
duke@0 2213 Node *n = call->in(TypeFunc::Parms);
duke@0 2214 int nop = n->Opcode();
duke@0 2215 // Clone shared simple arguments to uncommon calls, item (1).
duke@0 2216 if( n->outcnt() > 1 &&
duke@0 2217 !n->is_Proj() &&
duke@0 2218 nop != Op_CreateEx &&
duke@0 2219 nop != Op_CheckCastPP &&
kvn@331 2220 nop != Op_DecodeN &&
duke@0 2221 !n->is_Mem() ) {
duke@0 2222 Node *x = n->clone();
duke@0 2223 call->set_req( TypeFunc::Parms, x );
duke@0 2224 }
duke@0 2225 }
duke@0 2226 break;
duke@0 2227 }
duke@0 2228
duke@0 2229 case Op_StoreD:
duke@0 2230 case Op_LoadD:
duke@0 2231 case Op_LoadD_unaligned:
kvn@859 2232 frc.inc_double_count();
duke@0 2233 goto handle_mem;
duke@0 2234 case Op_StoreF:
duke@0 2235 case Op_LoadF:
kvn@859 2236 frc.inc_float_count();
duke@0 2237 goto handle_mem;
duke@0 2238
never@2345 2239 case Op_StoreCM:
never@2345 2240 {
never@2345 2241 // Convert OopStore dependence into precedence edge
never@2345 2242 Node* prec = n->in(MemNode::OopStore);
never@2345 2243 n->del_req(MemNode::OopStore);
never@2345 2244 n->add_prec(prec);
never@2345 2245 eliminate_redundant_card_marks(n);
never@2345 2246 }
never@2345 2247
never@2345 2248 // fall through
never@2345 2249
duke@0 2250 case Op_StoreB:
duke@0 2251 case Op_StoreC:
duke@0 2252 case Op_StorePConditional:
duke@0 2253 case Op_StoreI:
duke@0 2254 case Op_StoreL:
kvn@420 2255 case Op_StoreIConditional:
duke@0 2256 case Op_StoreLConditional:
duke@0 2257 case Op_CompareAndSwapI:
duke@0 2258 case Op_CompareAndSwapL:
duke@0 2259 case Op_CompareAndSwapP:
coleenp@113 2260 case Op_CompareAndSwapN:
duke@0 2261 case Op_StoreP:
coleenp@113 2262 case Op_StoreN:
duke@0 2263 case Op_LoadB:
twisti@624 2264 case Op_LoadUB:
twisti@558 2265 case Op_LoadUS:
duke@0 2266 case Op_LoadI:
twisti@624 2267 case Op_LoadUI2L:
duke@0 2268 case Op_LoadKlass:
kvn@164 2269 case Op_LoadNKlass:
duke@0 2270 case Op_LoadL:
duke@0 2271 case Op_LoadL_unaligned:
duke@0 2272 case Op_LoadPLocked:
duke@0 2273 case Op_LoadLLocked:
duke@0 2274 case Op_LoadP:
coleenp@113 2275 case Op_LoadN:
duke@0 2276 case Op_LoadRange:
duke@0 2277 case Op_LoadS: {
duke@0 2278 handle_mem:
duke@0 2279 #ifdef ASSERT
duke@0 2280 if( VerifyOptoOopOffsets ) {
duke@0 2281 assert( n->is_Mem(), "" );
duke@0 2282 MemNode *mem = (MemNode*)n;
duke@0 2283 // Check to see if address types have grounded out somehow.
duke@0 2284 const TypeInstPtr *tp = mem->in(MemNode::Address)->bottom_type()->isa_instptr();
duke@0 2285 assert( !tp || oop_offset_is_sane(tp), "" );
duke@0 2286 }
duke@0 2287 #endif
duke@0 2288 break;
duke@0 2289 }
duke@0 2290
duke@0 2291 case Op_AddP: { // Assert sane base pointers
kvn@182 2292 Node *addp = n->in(AddPNode::Address);
duke@0 2293 assert( !addp->is_AddP() ||
duke@0 2294 addp->in(AddPNode::Base)->is_top() || // Top OK for allocation
duke@0 2295 addp->in(AddPNode::Base) == n->in(AddPNode::Base),
duke@0 2296 "Base pointers must match" );
kvn@182 2297 #ifdef _LP64
kvn@182 2298 if (UseCompressedOops &&
kvn@182 2299 addp->Opcode() == Op_ConP &&
kvn@182 2300 addp == n->in(AddPNode::Base) &&
kvn@182 2301 n->in(AddPNode::Offset)->is_Con()) {
kvn@182 2302 // Use addressing with narrow klass to load with offset on x86.
kvn@182 2303 // On sparc loading 32-bits constant and decoding it have less
kvn@182 2304 // instructions (4) then load 64-bits constant (7).
kvn@182 2305 // Do this transformation here since IGVN will convert ConN back to ConP.
kvn@182 2306 const Type* t = addp->bottom_type();
kvn@182 2307 if (t->isa_oopptr()) {
kvn@182 2308 Node* nn = NULL;
kvn@182 2309
kvn@182 2310 // Look for existing ConN node of the same exact type.
kvn@182 2311 Compile* C = Compile::current();
kvn@182 2312 Node* r = C->root();
kvn@182 2313 uint cnt = r->outcnt();
kvn@182 2314 for (uint i = 0; i < cnt; i++) {
kvn@182 2315 Node* m = r->raw_out(i);
kvn@182 2316 if (m!= NULL && m->Opcode() == Op_ConN &&
kvn@221 2317 m->bottom_type()->make_ptr() == t) {
kvn@182 2318 nn = m;
kvn@182 2319 break;
kvn@182 2320 }
kvn@182 2321 }
kvn@182 2322 if (nn != NULL) {
kvn@182 2323 // Decode a narrow oop to match address
kvn@182 2324 // [R12 + narrow_oop_reg<<3 + offset]
kvn@182 2325 nn = new (C, 2) DecodeNNode(nn, t);
kvn@182 2326 n->set_req(AddPNode::Base, nn);
kvn@182 2327 n->set_req(AddPNode::Address, nn);
kvn@182 2328 if (addp->outcnt() == 0) {
kvn@182 2329 addp->disconnect_inputs(NULL);
kvn@182 2330 }
kvn@182 2331 }
kvn@182 2332 }
kvn@182 2333 }
kvn@182 2334 #endif
duke@0 2335 break;
duke@0 2336 }
duke@0 2337
kvn@164 2338 #ifdef _LP64
kvn@368 2339 case Op_CastPP:
kvn@1495 2340 if (n->in(1)->is_DecodeN() && Matcher::gen_narrow_oop_implicit_null_checks()) {
kvn@368 2341 Compile* C = Compile::current();
kvn@368 2342 Node* in1 = n->in(1);
kvn@368 2343 const Type* t = n->bottom_type();
kvn@368 2344 Node* new_in1 = in1->clone();
kvn@368 2345 new_in1->as_DecodeN()->set_type(t);
kvn@368 2346
kvn@1495 2347 if (!Matcher::narrow_oop_use_complex_address()) {
kvn@368 2348 //
kvn@368 2349 // x86, ARM and friends can handle 2 adds in addressing mode
kvn@368 2350 // and Matcher can fold a DecodeN node into address by using
kvn@368 2351 // a narrow oop directly and do implicit NULL check in address:
kvn@368 2352 //
kvn@368 2353 // [R12 + narrow_oop_reg<<3 + offset]
kvn@368 2354 // NullCheck narrow_oop_reg
kvn@368 2355 //
kvn@368 2356 // On other platforms (Sparc) we have to keep new DecodeN node and
kvn@368 2357 // use it to do implicit NULL check in address:
kvn@368 2358 //
kvn@368 2359 // decode_not_null narrow_oop_reg, base_reg
kvn@368 2360 // [base_reg + offset]
kvn@368 2361 // NullCheck base_reg
kvn@368 2362 //
twisti@605 2363 // Pin the new DecodeN node to non-null path on these platform (Sparc)
kvn@368 2364 // to keep the information to which NULL check the new DecodeN node
kvn@368 2365 // corresponds to use it as value in implicit_null_check().
kvn@368 2366 //
kvn@368 2367 new_in1->set_req(0, n->in(0));
kvn@368 2368 }
kvn@368 2369
kvn@368 2370 n->subsume_by(new_in1);
kvn@368 2371 if (in1->outcnt() == 0) {
kvn@368 2372 in1->disconnect_inputs(NULL);
kvn@368 2373 }
kvn@368 2374 }
kvn@368 2375 break;
kvn@368 2376
kvn@164 2377 case Op_CmpP:
kvn@168 2378 // Do this transformation here to preserve CmpPNode::sub() and
kvn@168 2379 // other TypePtr related Ideal optimizations (for example, ptr nullness).
kvn@331 2380 if (n->in(1)->is_DecodeN() || n->in(2)->is_DecodeN()) {
kvn@331 2381 Node* in1 = n->in(1);
kvn@331 2382 Node* in2 = n->in(2);
kvn@331 2383 if (!in1->is_DecodeN()) {
kvn@331 2384 in2 = in1;
kvn@331 2385 in1 = n->in(2);
kvn@331 2386 }
kvn@331 2387 assert(in1->is_DecodeN(), "sanity");
kvn@331 2388
kvn@164 2389 Compile* C = Compile::current();
kvn@331 2390 Node* new_in2 = NULL;
kvn@331 2391 if (in2->is_DecodeN()) {
kvn@331 2392 new_in2 = in2->in(1);
kvn@331 2393 } else if (in2->Opcode() == Op_ConP) {
kvn@331 2394 const Type* t = in2->bottom_type();
kvn@1495 2395 if (t == TypePtr::NULL_PTR) {
kvn@1495 2396 // Don't convert CmpP null check into CmpN if compressed
kvn@1495 2397 // oops implicit null check is not generated.
kvn@1495 2398 // This will allow to generate normal oop implicit null check.
kvn@1495 2399 if (Matcher::gen_narrow_oop_implicit_null_checks())
kvn@1495 2400 new_in2 = ConNode::make(C, TypeNarrowOop::NULL_PTR);
kvn@368 2401 //
kvn@368 2402 // This transformation together with CastPP transformation above
kvn@368 2403 // will generated code for implicit NULL checks for compressed oops.
kvn@368 2404 //
kvn@368 2405 // The original code after Optimize()
kvn@368 2406 //
kvn@368 2407 // LoadN memory, narrow_oop_reg
kvn@368 2408 // decode narrow_oop_reg, base_reg
kvn@368 2409 // CmpP base_reg, NULL
kvn@368 2410 // CastPP base_reg // NotNull
kvn@368 2411 // Load [base_reg + offset], val_reg
kvn@368 2412 //
kvn@368 2413 // after these transformations will be
kvn@368 2414 //
kvn@368 2415 // LoadN memory, narrow_oop_reg
kvn@368 2416 // CmpN narrow_oop_reg, NULL
kvn@368 2417 // decode_not_null narrow_oop_reg, base_reg
kvn@368 2418 // Load [base_reg + offset], val_reg
kvn@368 2419 //
kvn@368 2420 // and the uncommon path (== NULL) will use narrow_oop_reg directly
kvn@368 2421 // since narrow oops can be used in debug info now (see the code in
kvn@368 2422 // final_graph_reshaping_walk()).
kvn@368 2423 //
kvn@368 2424 // At the end the code will be matched to
kvn@368 2425 // on x86:
kvn@368 2426 //
kvn@368 2427 // Load_narrow_oop memory, narrow_oop_reg
kvn@368 2428 // Load [R12 + narrow_oop_reg<<3 + offset], val_reg
kvn@368 2429 // NullCheck narrow_oop_reg
kvn@368 2430 //
kvn@368 2431 // and on sparc:
kvn@368 2432 //
kvn@368 2433 // Load_narrow_oop memory, narrow_oop_reg
kvn@368 2434 // decode_not_null narrow_oop_reg, base_reg
kvn@368 2435 // Load [base_reg + offset], val_reg
kvn@368 2436 // NullCheck base_reg
kvn@368 2437 //
kvn@164 2438 } else if (t->isa_oopptr()) {
kvn@331 2439 new_in2 = ConNode::make(C, t->make_narrowoop());
kvn@164 2440 }
kvn@164 2441 }
kvn@331 2442 if (new_in2 != NULL) {
kvn@331 2443 Node* cmpN = new (C, 3) CmpNNode(in1->in(1), new_in2);
kvn@168 2444 n->subsume_by( cmpN );
kvn@331 2445 if (in1->outcnt() == 0) {
kvn@331 2446 in1->disconnect_inputs(NULL);
kvn@331 2447 }
kvn@331 2448 if (in2->outcnt() == 0) {
kvn@331 2449 in2->disconnect_inputs(NULL);
kvn@331 2450 }
kvn@164 2451 }
kvn@164 2452 }
kvn@293 2453 break;
kvn@368 2454
kvn@368 2455 case Op_DecodeN:
kvn@368 2456 assert(!n->in(1)->is_EncodeP(), "should be optimized out");
kvn@1495 2457 // DecodeN could be pinned when it can't be fold into
kvn@492 2458 // an address expression, see the code for Op_CastPP above.
kvn@1495 2459 assert(n->in(0) == NULL || !Matcher::narrow_oop_use_complex_address(), "no control");
kvn@368 2460 break;
kvn@368 2461
kvn@368 2462 case Op_EncodeP: {
kvn@368 2463 Node* in1 = n->in(1);
kvn@368 2464 if (in1->is_DecodeN()) {
kvn@368 2465 n->subsume_by(in1->in(1));
kvn@368 2466 } else if (in1->Opcode() == Op_ConP) {
kvn@368 2467 Compile* C = Compile::current();
kvn@368 2468 const Type* t = in1->bottom_type();
kvn@368 2469 if (t == TypePtr::NULL_PTR) {
kvn@368 2470 n->subsume_by(ConNode::make(C, TypeNarrowOop::NULL_PTR));
kvn@368 2471 } else if (t->isa_oopptr()) {
kvn@368 2472 n->subsume_by(ConNode::make(C, t->make_narrowoop()));
kvn@368 2473 }
kvn@368 2474 }
kvn@368 2475 if (in1->outcnt() == 0) {
kvn@368 2476 in1->disconnect_inputs(NULL);
kvn@368 2477 }
kvn@368 2478 break;
kvn@368 2479 }
kvn@368 2480
never@1080 2481 case Op_Proj: {
never@1080 2482 if (OptimizeStringConcat) {
never@1080 2483 ProjNode* p = n->as_Proj();
never@1080 2484 if (p->_is_io_use) {
never@1080 2485 // Separate projections were used for the exception path which
never@1080 2486 // are normally removed by a late inline. If it wasn't inlined
never@1080 2487 // then they will hang around and should just be replaced with
never@1080 2488 // the original one.
never@1080 2489 Node* proj = NULL;
never@1080 2490 // Replace with just one
never@1080 2491 for (SimpleDUIterator i(p->in(0)); i.has_next(); i.next()) {
never@1080 2492 Node *use = i.get();
never@1080 2493 if (use->is_Proj() && p != use && use->as_Proj()->_con == p->_con) {
never@1080 2494 proj = use;
never@1080 2495 break;
never@1080 2496 }
never@1080 2497 }
never@1080 2498 assert(p != NULL, "must be found");
never@1080 2499 p->subsume_by(proj);
never@1080 2500 }
never@1080 2501 }
never@1080 2502 break;
never@1080 2503 }
never@1080 2504
kvn@368 2505 case Op_Phi:
kvn@368 2506 if (n->as_Phi()->bottom_type()->isa_narrowoop()) {
kvn@368 2507 // The EncodeP optimization may create Phi with the same edges
kvn@368 2508 // for all paths. It is not handled well by Register Allocator.
kvn@368 2509 Node* unique_in = n->in(1);
kvn@368 2510 assert(unique_in != NULL, "");
kvn@368 2511 uint cnt = n->req();
kvn@368 2512 for (uint i = 2; i < cnt; i++) {
kvn@368 2513 Node* m = n->in(i);
kvn@368 2514 assert(m != NULL, "");
kvn@368 2515 if (unique_in != m)
kvn@368 2516 unique_in = NULL;
kvn@368 2517 }
kvn@368 2518 if (unique_in != NULL) {
kvn@368 2519 n->subsume_by(unique_in);
kvn@368 2520 }
kvn@368 2521 }
kvn@368 2522 break;
kvn@368 2523
kvn@164 2524 #endif
kvn@164 2525
duke@0 2526 case Op_ModI:
duke@0 2527 if (UseDivMod) {
duke@0 2528 // Check if a%b and a/b both exist
duke@0 2529 Node* d = n->find_similar(Op_DivI);
duke@0 2530 if (d) {
duke@0 2531 // Replace them with a fused divmod if supported
duke@0 2532 Compile* C = Compile::current();
duke@0 2533 if (Matcher::has_match_rule(Op_DivModI)) {
duke@0 2534 DivModINode* divmod = DivModINode::make(C, n);
kvn@168 2535 d->subsume_by(divmod->div_proj());
kvn@168 2536 n->subsume_by(divmod->mod_proj());
duke@0 2537 } else {
duke@0 2538 // replace a%b with a-((a/b)*b)
duke@0 2539 Node* mult = new (C, 3) MulINode(d, d->in(2));
duke@0 2540 Node* sub = new (C, 3) SubINode(d->in(1), mult);
kvn@168 2541 n->subsume_by( sub );
duke@0 2542 }
duke@0 2543 }
duke@0 2544 }
duke@0 2545 break;
duke@0 2546
duke@0 2547 case Op_ModL:
duke@0 2548 if (UseDivMod) {
duke@0 2549 // Check if a%b and a/b both exist
duke@0 2550 Node* d = n->find_similar(Op_DivL);
duke@0 2551 if (d) {
duke@0 2552 // Replace them with a fused divmod if supported
duke@0 2553 Compile* C = Compile::current();
duke@0 2554 if (Matcher::has_match_rule(Op_DivModL)) {
duke@0 2555 DivModLNode* divmod = DivModLNode::make(C, n);
kvn@168 2556 d->subsume_by(divmod->div_proj());
kvn@168 2557 n->subsume_by(divmod->mod_proj());
duke@0 2558 } else {
duke@0 2559 // replace a%b with a-((a/b)*b)
duke@0 2560 Node* mult = new (C, 3) MulLNode(d, d->in(2));
duke@0 2561 Node* sub = new (C, 3) SubLNode(d->in(1), mult);
kvn@168 2562 n->subsume_by( sub );
duke@0 2563 }
duke@0 2564 }
duke@0 2565 }
duke@0 2566 break;
duke@0 2567
duke@0 2568 case Op_Load16B:
duke@0 2569 case Op_Load8B:
duke@0 2570 case Op_Load4B:
duke@0 2571 case Op_Load8S:
duke@0 2572 case Op_Load4S:
duke@0 2573 case Op_Load2S:
duke@0 2574 case Op_Load8C:
duke@0 2575 case Op_Load4C:
duke@0 2576 case Op_Load2C:
duke@0 2577 case Op_Load4I:
duke@0 2578 case Op_Load2I:
duke@0 2579 case Op_Load2L:
duke@0 2580 case Op_Load4F:
duke@0 2581 case Op_Load2F:
duke@0 2582 case Op_Load2D:
duke@0 2583 case Op_Store16B:
duke@0 2584 case Op_Store8B:
duke@0 2585 case Op_Store4B:
duke@0 2586 case Op_Store8C:
duke@0 2587 case Op_Store4C:
duke@0 2588 case Op_Store2C:
duke@0 2589 case Op_Store4I:
duke@0 2590 case Op_Store2I:
duke@0 2591 case Op_Store2L:
duke@0 2592 case Op_Store4F:
duke@0 2593 case Op_Store2F:
duke@0 2594 case Op_Store2D:
duke@0 2595 break;
duke@0 2596
duke@0 2597 case Op_PackB:
duke@0 2598 case Op_PackS:
duke@0 2599 case Op_PackC:
duke@0 2600 case Op_PackI:
duke@0 2601 case Op_PackF:
duke@0 2602 case Op_PackL:
duke@0 2603 case Op_PackD:
duke@0 2604 if (n->req()-1 > 2) {
duke@0 2605 // Replace many operand PackNodes with a binary tree for matching
duke@0 2606 PackNode* p = (PackNode*) n;
duke@0 2607 Node* btp = p->binaryTreePack(Compile::current(), 1, n->req());
kvn@168 2608 n->subsume_by(btp);
duke@0 2609 }
duke@0 2610 break;
kvn@859 2611 case Op_Loop:
kvn@859 2612 case Op_CountedLoop:
kvn@859 2613 if (n->as_Loop()->is_inner_loop()) {
kvn@859 2614 frc.inc_inner_loop_count();
kvn@859 2615 }
kvn@859 2616 break;
roland@2248 2617 case Op_LShiftI:
roland@2248 2618 case Op_RShiftI:
roland@2248 2619 case Op_URShiftI:
roland@2248 2620 case Op_LShiftL:
roland@2248 2621 case Op_RShiftL:
roland@2248 2622 case Op_URShiftL:
roland@2248 2623 if (Matcher::need_masked_shift_count) {
roland@2248 2624 // The cpu's shift instructions don't restrict the count to the
roland@2248 2625 // lower 5/6 bits. We need to do the masking ourselves.
roland@2248 2626 Node* in2 = n->in(2);
roland@2248 2627 juint mask = (n->bottom_type() == TypeInt::INT) ? (BitsPerInt - 1) : (BitsPerLong - 1);
roland@2248 2628 const TypeInt* t = in2->find_int_type();
roland@2248 2629 if (t != NULL && t->is_con()) {
roland@2248 2630 juint shift = t->get_con();
roland@2248 2631 if (shift > mask) { // Unsigned cmp
roland@2248 2632 Compile* C = Compile::current();
roland@2248 2633 n->set_req(2, ConNode::make(C, TypeInt::make(shift & mask)));
roland@2248 2634 }
roland@2248 2635 } else {
roland@2248 2636 if (t == NULL || t->_lo < 0 || t->_hi > (int)mask) {
roland@2248 2637 Compile* C = Compile::current();
roland@2248 2638 Node* shift = new (C, 3) AndINode(in2, ConNode::make(C, TypeInt::make(mask)));
roland@2248 2639 n->set_req(2, shift);
roland@2248 2640 }
roland@2248 2641 }
roland@2248 2642 if (in2->outcnt() == 0) { // Remove dead node
roland@2248 2643 in2->disconnect_inputs(NULL);
roland@2248 2644 }
roland@2248 2645 }
roland@2248 2646 break;
duke@0 2647 default:
duke@0 2648 assert( !n->is_Call(), "" );
duke@0 2649 assert( !n->is_Mem(), "" );
duke@0 2650 break;
duke@0 2651 }
never@127 2652
never@127 2653 // Collect CFG split points
never@127 2654 if (n->is_MultiBranch())
kvn@859 2655 frc._tests.push(n);
duke@0 2656 }
duke@0 2657
duke@0 2658 //------------------------------final_graph_reshaping_walk---------------------
duke@0 2659 // Replacing Opaque nodes with their input in final_graph_reshaping_impl(),
duke@0 2660 // requires that the walk visits a node's inputs before visiting the node.
kvn@859 2661 static void final_graph_reshaping_walk( Node_Stack &nstack, Node *root, Final_Reshape_Counts &frc ) {
kvn@331 2662 ResourceArea *area = Thread::current()->resource_area();
kvn@331 2663 Unique_Node_List sfpt(area);
kvn@331 2664
kvn@859 2665 frc._visited.set(root->_idx); // first, mark node as visited
duke@0 2666 uint cnt = root->req();
duke@0 2667 Node *n = root;
duke@0 2668 uint i = 0;
duke@0 2669 while (true) {
duke@0 2670 if (i < cnt) {
duke@0 2671 // Place all non-visited non-null inputs onto stack
duke@0 2672 Node* m = n->in(i);
duke@0 2673 ++i;
kvn@859 2674 if (m != NULL && !frc._visited.test_set(m->_idx)) {
kvn@331 2675 if (m->is_SafePoint() && m->as_SafePoint()->jvms() != NULL)
kvn@331 2676 sfpt.push(m);
duke@0 2677 cnt = m->req();
duke@0 2678 nstack.push(n, i); // put on stack parent and next input's index
duke@0 2679 n = m;
duke@0 2680 i = 0;
duke@0 2681 }
duke@0 2682 } else {
duke@0 2683 // Now do post-visit work
kvn@859 2684 final_graph_reshaping_impl( n, frc );
duke@0 2685 if (nstack.is_empty())
duke@0 2686 break; // finished
duke@0 2687 n = nstack.node(); // Get node from stack
duke@0 2688 cnt = n->req();
duke@0 2689 i = nstack.index();
duke@0 2690 nstack.pop(); // Shift to the next node on stack
duke@0 2691 }
duke@0 2692 }
kvn@331 2693
kvn@1495 2694 // Skip next transformation if compressed oops are not used.
kvn@1495 2695 if (!UseCompressedOops || !Matcher::gen_narrow_oop_implicit_null_checks())
kvn@1495 2696 return;
kvn@1495 2697
kvn@331 2698 // Go over safepoints nodes to skip DecodeN nodes for debug edges.
kvn@331 2699 // It could be done for an uncommon traps or any safepoints/calls
kvn@331 2700 // if the DecodeN node is referenced only in a debug info.
kvn@331 2701 while (sfpt.size() > 0) {
kvn@331 2702 n = sfpt.pop();
kvn@331 2703 JVMState *jvms = n->as_SafePoint()->jvms();
kvn@331 2704 assert(jvms != NULL, "sanity");
kvn@331 2705 int start = jvms->debug_start();
kvn@331 2706 int end = n->req();
kvn@331 2707 bool is_uncommon = (n->is_CallStaticJava() &&
kvn@331 2708 n->as_CallStaticJava()->uncommon_trap_request() != 0);
kvn@331 2709 for (int j = start; j < end; j++) {
kvn@331 2710 Node* in = n->in(j);
kvn@331 2711 if (in->is_DecodeN()) {
kvn@331 2712 bool safe_to_skip = true;
kvn@331 2713 if (!is_uncommon ) {
kvn@331 2714 // Is it safe to skip?
kvn@331 2715 for (uint i = 0; i < in->outcnt(); i++) {
kvn@331 2716 Node* u = in->raw_out(i);
kvn@331 2717 if (!u->is_SafePoint() ||
kvn@331 2718 u->is_Call() && u->as_Call()->has_non_debug_use(n)) {
kvn@331 2719 safe_to_skip = false;
kvn@331 2720 }
kvn@331 2721 }
kvn@331 2722 }
kvn@331 2723 if (safe_to_skip) {
kvn@331 2724 n->set_req(j, in->in(1));
kvn@331 2725 }
kvn@331 2726 if (in->outcnt() == 0) {
kvn@331 2727 in->disconnect_inputs(NULL);
kvn@331 2728 }
kvn@331 2729 }
kvn@331 2730 }
kvn@331 2731 }
duke@0 2732 }
duke@0 2733
duke@0 2734 //------------------------------final_graph_reshaping--------------------------
duke@0 2735 // Final Graph Reshaping.
duke@0 2736 //
duke@0 2737 // (1) Clone simple inputs to uncommon calls, so they can be scheduled late
duke@0 2738 // and not commoned up and forced early. Must come after regular
duke@0 2739 // optimizations to avoid GVN undoing the cloning. Clone constant
duke@0 2740 // inputs to Loop Phis; these will be split by the allocator anyways.
duke@0 2741 // Remove Opaque nodes.
duke@0 2742 // (2) Move last-uses by commutative operations to the left input to encourage
duke@0 2743 // Intel update-in-place two-address operations and better register usage
duke@0 2744 // on RISCs. Must come after regular optimizations to avoid GVN Ideal
duke@0 2745 // calls canonicalizing them back.
duke@0 2746 // (3) Count the number of double-precision FP ops, single-precision FP ops
duke@0 2747 // and call sites. On Intel, we can get correct rounding either by
duke@0 2748 // forcing singles to memory (requires extra stores and loads after each
duke@0 2749 // FP bytecode) or we can set a rounding mode bit (requires setting and
duke@0 2750 // clearing the mode bit around call sites). The mode bit is only used
duke@0 2751 // if the relative frequency of single FP ops to calls is low enough.
duke@0 2752 // This is a key transform for SPEC mpeg_audio.
duke@0 2753 // (4) Detect infinite loops; blobs of code reachable from above but not
duke@0 2754 // below. Several of the Code_Gen algorithms fail on such code shapes,
duke@0 2755 // so we simply bail out. Happens a lot in ZKM.jar, but also happens
duke@0 2756 // from time to time in other codes (such as -Xcomp finalizer loops, etc).
duke@0 2757 // Detection is by looking for IfNodes where only 1 projection is
duke@0 2758 // reachable from below or CatchNodes missing some targets.
duke@0 2759 // (5) Assert for insane oop offsets in debug mode.
duke@0 2760
duke@0 2761 bool Compile::final_graph_reshaping() {
duke@0 2762 // an infinite loop may have been eliminated by the optimizer,
duke@0 2763 // in which case the graph will be empty.
duke@0 2764 if (root()->req() == 1) {
duke@0 2765 record_method_not_compilable("trivial infinite loop");
duke@0 2766 return true;
duke@0 2767 }
duke@0 2768
kvn@859 2769 Final_Reshape_Counts frc;
duke@0 2770
duke@0 2771 // Visit everybody reachable!
duke@0 2772 // Allocate stack of size C->unique()/2 to avoid frequent realloc
duke@0 2773 Node_Stack nstack(unique() >> 1);
kvn@859 2774 final_graph_reshaping_walk(nstack, root(), frc);
duke@0 2775
duke@0 2776 // Check for unreachable (from below) code (i.e., infinite loops).
kvn@859 2777 for( uint i = 0; i < frc._tests.size(); i++ ) {
kvn@859 2778 MultiBranchNode *n = frc._tests[i]->as_MultiBranch();
never@127 2779 // Get number of CFG targets.
duke@0 2780 // Note that PCTables include exception targets after calls.
never@127 2781 uint required_outcnt = n->required_outcnt();
never@127 2782 if (n->outcnt() != required_outcnt) {
duke@0 2783 // Check for a few special cases. Rethrow Nodes never take the
duke@0 2784 // 'fall-thru' path, so expected kids is 1 less.
duke@0 2785 if (n->is_PCTable() && n->in(0) && n->in(0)->in(0)) {
duke@0 2786 if (n->in(0)->in(0)->is_Call()) {
duke@0 2787 CallNode *call = n->in(0)->in(0)->as_Call();
duke@0 2788 if (call->entry_point() == OptoRuntime::rethrow_stub()) {
never@127 2789 required_outcnt--; // Rethrow always has 1 less kid
duke@0 2790 } else if (call->req() > TypeFunc::Parms &&
duke@0 2791 call->is_CallDynamicJava()) {
duke@0 2792 // Check for null receiver. In such case, the optimizer has
duke@0 2793 // detected that the virtual call will always result in a null
duke@0 2794 // pointer exception. The fall-through projection of this CatchNode
duke@0 2795 // will not be populated.
duke@0 2796 Node *arg0 = call->in(TypeFunc::Parms);
duke@0 2797 if (arg0->is_Type() &&
duke@0 2798 arg0->as_Type()->type()->higher_equal(TypePtr::NULL_PTR)) {
never@127 2799 required_outcnt--;
duke@0 2800 }
duke@0 2801 } else if (call->entry_point() == OptoRuntime::new_array_Java() &&
duke@0 2802 call->req() > TypeFunc::Parms+1 &&
duke@0 2803 call->is_CallStaticJava()) {
duke@0 2804 // Check for negative array length. In such case, the optimizer has
duke@0 2805 // detected that the allocation attempt will always result in an
duke@0 2806 // exception. There is no fall-through projection of this CatchNode .
duke@0 2807 Node *arg1 = call->in(TypeFunc::Parms+1);
duke@0 2808 if (arg1->is_Type() &&
duke@0 2809 arg1->as_Type()->type()->join(TypeInt::POS)->empty()) {
never@127 2810 required_outcnt--;
duke@0 2811 }
duke@0 2812 }
duke@0 2813 }
duke@0 2814 }
never@127 2815 // Recheck with a better notion of 'required_outcnt'
never@127 2816 if (n->outcnt() != required_outcnt) {
duke@0 2817 record_method_not_compilable("malformed control flow");
duke@0 2818 return true; // Not all targets reachable!
duke@0 2819 }
duke@0 2820 }
duke@0 2821 // Check that I actually visited all kids. Unreached kids
duke@0 2822 // must be infinite loops.
duke@0 2823 for (DUIterator_Fast jmax, j = n->fast_outs(jmax); j < jmax; j++)
kvn@859 2824 if (!frc._visited.test(n->fast_out(j)->_idx)) {
duke@0 2825 record_method_not_compilable("infinite loop");
duke@0 2826 return true; // Found unvisited kid; must be unreach
duke@0 2827 }
duke@0 2828 }
duke@0 2829
duke@0 2830 // If original bytecodes contained a mixture of floats and doubles
duke@0 2831 // check if the optimizer has made it homogenous, item (3).
never@929 2832 if( Use24BitFPMode && Use24BitFP && UseSSE == 0 &&
kvn@859 2833 frc.get_float_count() > 32 &&
kvn@859 2834 frc.get_double_count() == 0 &&
kvn@859 2835 (10 * frc.get_call_count() < frc.get_float_count()) ) {
duke@0 2836 set_24_bit_selection_and_mode( false, true );
duke@0 2837 }
duke@0 2838
kvn@859 2839 set_java_calls(frc.get_java_call_count());
kvn@859 2840 set_inner_loops(frc.get_inner_loop_count());
duke@0 2841
duke@0 2842 // No infinite loops, no reason to bail out.
duke@0 2843 return false;
duke@0 2844 }
duke@0 2845
duke@0 2846 //-----------------------------too_many_traps----------------------------------
duke@0 2847 // Report if there are too many traps at the current method and bci.
duke@0 2848 // Return true if there was a trap, and/or PerMethodTrapLimit is exceeded.
duke@0 2849 bool Compile::too_many_traps(ciMethod* method,
duke@0 2850 int bci,
duke@0 2851 Deoptimization::DeoptReason reason) {
duke@0 2852 ciMethodData* md = method->method_data();
duke@0 2853 if (md->is_empty()) {
duke@0 2854 // Assume the trap has not occurred, or that it occurred only
duke@0 2855 // because of a transient condition during start-up in the interpreter.
duke@0 2856 return false;
duke@0 2857 }
duke@0 2858 if (md->has_trap_at(bci, reason) != 0) {
duke@0 2859 // Assume PerBytecodeTrapLimit==0, for a more conservative heuristic.
duke@0 2860 // Also, if there are multiple reasons, or if there is no per-BCI record,
duke@0 2861 // assume the worst.
duke@0 2862 if (log())
duke@0 2863 log()->elem("observe trap='%s' count='%d'",
duke@0 2864 Deoptimization::trap_reason_name(reason),
duke@0 2865 md->trap_count(reason));
duke@0 2866 return true;
duke@0 2867 } else {
duke@0 2868 // Ignore method/bci and see if there have been too many globally.
duke@0 2869 return too_many_traps(reason, md);
duke@0 2870 }
duke@0 2871 }
duke@0 2872
duke@0 2873 // Less-accurate variant which does not require a method and bci.
duke@0 2874 bool Compile::too_many_traps(Deoptimization::DeoptReason reason,
duke@0 2875 ciMethodData* logmd) {
duke@0 2876 if (trap_count(reason) >= (uint)PerMethodTrapLimit) {
duke@0 2877 // Too many traps globally.
duke@0 2878 // Note that we use cumulative trap_count, not just md->trap_count.
duke@0 2879 if (log()) {
duke@0 2880 int mcount = (logmd == NULL)? -1: (int)logmd->trap_count(reason);
duke@0 2881 log()->elem("observe trap='%s' count='0' mcount='%d' ccount='%d'",
duke@0 2882 Deoptimization::trap_reason_name(reason),
duke@0 2883 mcount, trap_count(reason));
duke@0 2884 }
duke@0 2885 return true;
duke@0 2886 } else {
duke@0 2887 // The coast is clear.
duke@0 2888 return false;
duke@0 2889 }
duke@0 2890 }
duke@0 2891
duke@0 2892 //--------------------------too_many_recompiles--------------------------------
duke@0 2893 // Report if there are too many recompiles at the current method and bci.
duke@0 2894 // Consults PerBytecodeRecompilationCutoff and PerMethodRecompilationCutoff.
duke@0 2895 // Is not eager to return true, since this will cause the compiler to use
duke@0 2896 // Action_none for a trap point, to avoid too many recompilations.
duke@0 2897 bool Compile::too_many_recompiles(ciMethod* method,
duke@0 2898 int bci,
duke@0 2899 Deoptimization::DeoptReason reason) {
duke@0 2900 ciMethodData* md = method->method_data();
duke@0 2901 if (md->is_empty()) {
duke@0 2902 // Assume the trap has not occurred, or that it occurred only
duke@0 2903 // because of a transient condition during start-up in the interpreter.
duke@0 2904 return false;
duke@0 2905 }
duke@0 2906 // Pick a cutoff point well within PerBytecodeRecompilationCutoff.
duke@0 2907 uint bc_cutoff = (uint) PerBytecodeRecompilationCutoff / 8;
duke@0 2908 uint m_cutoff = (uint) PerMethodRecompilationCutoff / 2 + 1; // not zero
duke@0 2909 Deoptimization::DeoptReason per_bc_reason
duke@0 2910 = Deoptimization::reason_recorded_per_bytecode_if_any(reason);
duke@0 2911 if ((per_bc_reason == Deoptimization::Reason_none
duke@0 2912 || md->has_trap_at(bci, reason) != 0)
duke@0 2913 // The trap frequency measure we care about is the recompile count:
duke@0 2914 && md->trap_recompiled_at(bci)
duke@0 2915 && md->overflow_recompile_count() >= bc_cutoff) {
duke@0 2916 // Do not emit a trap here if it has already caused recompilations.
duke@0 2917 // Also, if there are multiple reasons, or if there is no per-BCI record,
duke@0 2918 // assume the worst.
duke@0 2919 if (log())
duke@0 2920 log()->elem("observe trap='%s recompiled' count='%d' recompiles2='%d'",
duke@0 2921 Deoptimization::trap_reason_name(reason),
duke@0 2922 md->trap_count(reason),
duke@0 2923 md->overflow_recompile_count());
duke@0 2924 return true;
duke@0 2925 } else if (trap_count(reason) != 0
duke@0 2926 && decompile_count() >= m_cutoff) {
duke@0 2927 // Too many recompiles globally, and we have seen this sort of trap.
duke@0 2928 // Use cumulative decompile_count, not just md->decompile_count.
duke@0 2929 if (log())
duke@0 2930 log()->elem("observe trap='%s' count='%d' mcount='%d' decompiles='%d' mdecompiles='%d'",
duke@0 2931 Deoptimization::trap_reason_name(reason),
duke@0 2932 md->trap_count(reason), trap_count(reason),
duke@0 2933 md->decompile_count(), decompile_count());
duke@0 2934 return true;
duke@0 2935 } else {
duke@0 2936 // The coast is clear.
duke@0 2937 return false;
duke@0 2938 }
duke@0 2939 }
duke@0 2940
duke@0 2941
duke@0 2942 #ifndef PRODUCT
duke@0 2943 //------------------------------verify_graph_edges---------------------------
duke@0 2944 // Walk the Graph and verify that there is a one-to-one correspondence
duke@0 2945 // between Use-Def edges and Def-Use edges in the graph.
duke@0 2946 void Compile::verify_graph_edges(bool no_dead_code) {
duke@0 2947 if (VerifyGraphEdges) {
duke@0 2948 ResourceArea *area = Thread::current()->resource_area();
duke@0 2949 Unique_Node_List visited(area);
duke@0 2950 // Call recursive graph walk to check edges
duke@0 2951 _root->verify_edges(visited);
duke@0 2952 if (no_dead_code) {
duke@0 2953 // Now make sure that no visited node is used by an unvisited node.
duke@0 2954 bool dead_nodes = 0;
duke@0 2955 Unique_Node_List checked(area);
duke@0 2956 while (visited.size() > 0) {
duke@0 2957 Node* n = visited.pop();
duke@0 2958 checked.push(n);
duke@0 2959 for (uint i = 0; i < n->outcnt(); i++) {
duke@0 2960 Node* use = n->raw_out(i);
duke@0 2961 if (checked.member(use)) continue; // already checked
duke@0 2962 if (visited.member(use)) continue; // already in the graph
duke@0 2963 if (use->is_Con()) continue; // a dead ConNode is OK
duke@0 2964 // At this point, we have found a dead node which is DU-reachable.
duke@0 2965 if (dead_nodes++ == 0)
duke@0 2966 tty->print_cr("*** Dead nodes reachable via DU edges:");
duke@0 2967 use->dump(2);
duke@0 2968 tty->print_cr("---");
duke@0 2969 checked.push(use); // No repeats; pretend it is now checked.
duke@0 2970 }
duke@0 2971 }
duke@0 2972 assert(dead_nodes == 0, "using nodes must be reachable from root");
duke@0 2973 }
duke@0 2974 }
duke@0 2975 }
duke@0 2976 #endif
duke@0 2977
duke@0 2978 // The Compile object keeps track of failure reasons separately from the ciEnv.
duke@0 2979 // This is required because there is not quite a 1-1 relation between the
duke@0 2980 // ciEnv and its compilation task and the Compile object. Note that one
duke@0 2981 // ciEnv might use two Compile objects, if C2Compiler::compile_method decides
duke@0 2982 // to backtrack and retry without subsuming loads. Other than this backtracking
duke@0 2983 // behavior, the Compile's failure reason is quietly copied up to the ciEnv
duke@0 2984 // by the logic in C2Compiler.
duke@0 2985 void Compile::record_failure(const char* reason) {
duke@0 2986 if (log() != NULL) {
duke@0 2987 log()->elem("failure reason='%s' phase='compile'", reason);
duke@0 2988 }
duke@0 2989 if (_failure_reason == NULL) {
duke@0 2990 // Record the first failure reason.
duke@0 2991 _failure_reason = reason;
duke@0 2992 }
never@222 2993 if (!C->failure_reason_is(C2Compiler::retry_no_subsuming_loads())) {
never@222 2994 C->print_method(_failure_reason);
never@222 2995 }
duke@0 2996 _root = NULL; // flush the graph, too
duke@0 2997 }
duke@0 2998
duke@0 2999 Compile::TracePhase::TracePhase(const char* name, elapsedTimer* accumulator, bool dolog)
duke@0 3000 : TraceTime(NULL, accumulator, false NOT_PRODUCT( || TimeCompiler ), false)
duke@0 3001 {
duke@0 3002 if (dolog) {
duke@0 3003 C = Compile::current();
duke@0 3004 _log = C->log();
duke@0 3005 } else {
duke@0 3006 C = NULL;
duke@0 3007 _log = NULL;
duke@0 3008 }
duke@0 3009 if (_log != NULL) {
duke@0 3010 _log->begin_head("phase name='%s' nodes='%d'", name, C->unique());
duke@0 3011 _log->stamp();
duke@0 3012 _log->end_head();
duke@0 3013 }
duke@0 3014 }
duke@0 3015
duke@0 3016 Compile::TracePhase::~TracePhase() {
duke@0 3017 if (_log != NULL) {
duke@0 3018 _log->done("phase nodes='%d'", C->unique());
duke@0 3019 }
duke@0 3020 }
twisti@1915 3021
twisti@1915 3022 //=============================================================================
twisti@1915 3023 // Two Constant's are equal when the type and the value are equal.
twisti@1915 3024 bool Compile::Constant::operator==(const Constant& other) {
twisti@1915 3025 if (type() != other.type() ) return false;
twisti@1915 3026 if (can_be_reused() != other.can_be_reused()) return false;
twisti@1915 3027 // For floating point values we compare the bit pattern.
twisti@1915 3028 switch (type()) {
twisti@1915 3029 case T_FLOAT: return (_value.i == other._value.i);
twisti@1915 3030 case T_LONG:
twisti@1915 3031 case T_DOUBLE: return (_value.j == other._value.j);
twisti@1915 3032 case T_OBJECT:
twisti@1915 3033 case T_ADDRESS: return (_value.l == other._value.l);
twisti@1915 3034 case T_VOID: return (_value.l == other._value.l); // jump-table entries
twisti@1915 3035 default: ShouldNotReachHere();
twisti@1915 3036 }
twisti@1915 3037 return false;
twisti@1915 3038 }
twisti@1915 3039
twisti@1915 3040 // Emit constants grouped in the following order:
twisti@1915 3041 static BasicType type_order[] = {
twisti@1915 3042 T_FLOAT, // 32-bit
twisti@1915 3043 T_OBJECT, // 32 or 64-bit
twisti@1915 3044 T_ADDRESS, // 32 or 64-bit
twisti@1915 3045 T_DOUBLE, // 64-bit
twisti@1915 3046 T_LONG, // 64-bit
twisti@1915 3047 T_VOID, // 32 or 64-bit (jump-tables are at the end of the constant table for code emission reasons)
twisti@1915 3048 T_ILLEGAL
twisti@1915 3049 };
twisti@1915 3050
twisti@1915 3051 static int type_to_size_in_bytes(BasicType t) {
twisti@1915 3052 switch (t) {
twisti@1915 3053 case T_LONG: return sizeof(jlong );
twisti@1915 3054 case T_FLOAT: return sizeof(jfloat );
twisti@1915 3055 case T_DOUBLE: return sizeof(jdouble);
twisti@1915 3056 // We use T_VOID as marker for jump-table entries (labels) which
twisti@1915 3057 // need an interal word relocation.
twisti@1915 3058 case T_VOID:
twisti@1915 3059 case T_ADDRESS:
twisti@1915 3060 case T_OBJECT: return sizeof(jobject);
twisti@1915 3061 }
twisti@1915 3062
twisti@1915 3063 ShouldNotReachHere();
twisti@1915 3064 return -1;
twisti@1915 3065 }
twisti@1915 3066
twisti@1915 3067 void Compile::ConstantTable::calculate_offsets_and_size() {
twisti@1915 3068 int size = 0;
twisti@1915 3069 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
twisti@1915 3070 BasicType type = type_order[t];
twisti@1915 3071
twisti@1915 3072 for (int i = 0; i < _constants.length(); i++) {
twisti@1915 3073 Constant con = _constants.at(i);
twisti@1915 3074 if (con.type() != type) continue; // Skip other types.
twisti@1915 3075
twisti@1915 3076 // Align size for type.
twisti@1915 3077 int typesize = type_to_size_in_bytes(con.type());
twisti@1915 3078 size = align_size_up(size, typesize);
twisti@1915 3079
twisti@1915 3080 // Set offset.
twisti@1915 3081 con.set_offset(size);
twisti@1915 3082 _constants.at_put(i, con);
twisti@1915 3083
twisti@1915 3084 // Add type size.
twisti@1915 3085 size = size + typesize;
twisti@1915 3086 }
twisti@1915 3087 }
twisti@1915 3088
twisti@1915 3089 // Align size up to the next section start (which is insts; see
twisti@1915 3090 // CodeBuffer::align_at_start).
twisti@1915 3091 assert(_size == -1, "already set?");
twisti@1915 3092 _size = align_size_up(size, CodeEntryAlignment);
twisti@1915 3093
twisti@1915 3094 if (Matcher::constant_table_absolute_addressing) {
twisti@1915 3095 set_table_base_offset(0); // No table base offset required
twisti@1915 3096 } else {
twisti@1915 3097 if (UseRDPCForConstantTableBase) {
twisti@1915 3098 // table base offset is set in MachConstantBaseNode::emit
twisti@1915 3099 } else {
twisti@1915 3100 // When RDPC is not used, the table base is set into the middle of
twisti@1915 3101 // the constant table.
twisti@1915 3102 int half_size = _size / 2;
twisti@1915 3103 assert(half_size * 2 == _size, "sanity");
twisti@1915 3104 set_table_base_offset(-half_size);
twisti@1915 3105 }
twisti@1915 3106 }
twisti@1915 3107 }
twisti@1915 3108
twisti@1915 3109 void Compile::ConstantTable::emit(CodeBuffer& cb) {
twisti@1915 3110 MacroAssembler _masm(&cb);
twisti@1915 3111 for (int t = 0; type_order[t] != T_ILLEGAL; t++) {
twisti@1915 3112 BasicType type = type_order[t];
twisti@1915 3113
twisti@1915 3114 for (int i = 0; i < _constants.length(); i++) {
twisti@1915 3115 Constant con = _constants.at(i);
twisti@1915 3116 if (con.type() != type) continue; // Skip other types.
twisti@1915 3117
twisti@1915 3118 address constant_addr;
twisti@1915 3119 switch (con.type()) {
twisti@1915 3120 case T_LONG: constant_addr = _masm.long_constant( con.get_jlong() ); break;
twisti@1915 3121 case T_FLOAT: constant_addr = _masm.float_constant( con.get_jfloat() ); break;
twisti@1915 3122 case T_DOUBLE: constant_addr = _masm.double_constant(con.get_jdouble()); break;
twisti@1915 3123 case T_OBJECT: {
twisti@1915 3124 jobject obj = con.get_jobject();
twisti@1915 3125 int oop_index = _masm.oop_recorder()->find_index(obj);
twisti@1915 3126 constant_addr = _masm.address_constant((address) obj, oop_Relocation::spec(oop_index));
twisti@1915 3127 break;
twisti@1915 3128 }
twisti@1915 3129 case T_ADDRESS: {
twisti@1915 3130 address addr = (address) con.get_jobject();
twisti@1915 3131 constant_addr = _masm.address_constant(addr);
twisti@1915 3132 break;
twisti@1915 3133 }
twisti@1915 3134 // We use T_VOID as marker for jump-table entries (labels) which
twisti@1915 3135 // need an interal word relocation.
twisti@1915 3136 case T_VOID: {
twisti@1915 3137 // Write a dummy word. The real value is filled in later
twisti@1915 3138 // in fill_jump_table_in_constant_table.
twisti@1915 3139 address addr = (address) con.get_jobject();
twisti@1915 3140 constant_addr = _masm.address_constant(addr);
twisti@1915 3141 break;
twisti@1915 3142 }
twisti@1915 3143 default: ShouldNotReachHere();
twisti@1915 3144 }
twisti@1915 3145 assert(constant_addr != NULL, "consts section too small");
twisti@1915 3146 assert((constant_addr - _masm.code()->consts()->start()) == con.offset(), err_msg("must be: %d == %d", constant_addr - _masm.code()->consts()->start(), con.offset()));
twisti@1915 3147 }
twisti@1915 3148 }
twisti@1915 3149 }
twisti@1915 3150
twisti@1915 3151 int Compile::ConstantTable::find_offset(Constant& con) const {
twisti@1915 3152 int idx = _constants.find(con);
twisti@1915 3153 assert(idx != -1, "constant must be in constant table");
twisti@1915 3154 int offset = _constants.at(idx).offset();
twisti@1915 3155 assert(offset != -1, "constant table not emitted yet?");
twisti@1915 3156 return offset;
twisti@1915 3157 }
twisti@1915 3158
twisti@1915 3159 void Compile::ConstantTable::add(Constant& con) {
twisti@1915 3160 if (con.can_be_reused()) {
twisti@1915 3161 int idx = _constants.find(con);
twisti@1915 3162 if (idx != -1 && _constants.at(idx).can_be_reused()) {
twisti@1915 3163 return;
twisti@1915 3164 }
twisti@1915 3165 }
twisti@1915 3166 (void) _constants.append(con);
twisti@1915 3167 }
twisti@1915 3168
twisti@1915 3169 Compile::Constant Compile::ConstantTable::add(BasicType type, jvalue value) {
twisti@1915 3170 Constant con(type, value);
twisti@1915 3171 add(con);
twisti@1915 3172 return con;
twisti@1915 3173 }
twisti@1915 3174
twisti@1915 3175 Compile::Constant Compile::ConstantTable::add(MachOper* oper) {
twisti@1915 3176 jvalue value;
twisti@1915 3177 BasicType type = oper->type()->basic_type();
twisti@1915 3178 switch (type) {
twisti@1915 3179 case T_LONG: value.j = oper->constantL(); break;
twisti@1915 3180 case T_FLOAT: value.f = oper->constantF(); break;
twisti@1915 3181 case T_DOUBLE: value.d = oper->constantD(); break;
twisti@1915 3182 case T_OBJECT:
twisti@1915 3183 case T_ADDRESS: value.l = (jobject) oper->constant(); break;
twisti@1915 3184 default: ShouldNotReachHere();
twisti@1915 3185 }
twisti@1915 3186 return add(type, value);
twisti@1915 3187 }
twisti@1915 3188
twisti@1915 3189 Compile::Constant Compile::ConstantTable::allocate_jump_table(MachConstantNode* n) {
twisti@1915 3190 jvalue value;
twisti@1915 3191 // We can use the node pointer here to identify the right jump-table
twisti@1915 3192 // as this method is called from Compile::Fill_buffer right before
twisti@1915 3193 // the MachNodes are emitted and the jump-table is filled (means the
twisti@1915 3194 // MachNode pointers do not change anymore).
twisti@1915 3195 value.l = (jobject) n;
twisti@1915 3196 Constant con(T_VOID, value, false); // Labels of a jump-table cannot be reused.
twisti@1915 3197 for (uint i = 0; i < n->outcnt(); i++) {
twisti@1915 3198 add(con);
twisti@1915 3199 }
twisti@1915 3200 return con;
twisti@1915 3201 }
twisti@1915 3202
twisti@1915 3203 void Compile::ConstantTable::fill_jump_table(CodeBuffer& cb, MachConstantNode* n, GrowableArray<Label*> labels) const {
twisti@1915 3204 // If called from Compile::scratch_emit_size do nothing.
twisti@1915 3205 if (Compile::current()->in_scratch_emit_size()) return;
twisti@1915 3206
twisti@1915 3207 assert(labels.is_nonempty(), "must be");
twisti@1915 3208 assert((uint) labels.length() == n->outcnt(), err_msg("must be equal: %d == %d", labels.length(), n->outcnt()));
twisti@1915 3209
twisti@1915 3210 // Since MachConstantNode::constant_offset() also contains
twisti@1915 3211 // table_base_offset() we need to subtract the table_base_offset()
twisti@1915 3212 // to get the plain offset into the constant table.
twisti@1915 3213 int offset = n->constant_offset() - table_base_offset();
twisti@1915 3214
twisti@1915 3215 MacroAssembler _masm(&cb);
twisti@1915 3216 address* jump_table_base = (address*) (_masm.code()->consts()->start() + offset);
twisti@1915 3217
twisti@1915 3218 for (int i = 0; i < labels.length(); i++) {
twisti@1915 3219 address* constant_addr = &jump_table_base[i];
twisti@1915 3220 assert(*constant_addr == (address) n, "all jump-table entries must contain node pointer");
twisti@1915 3221 *constant_addr = cb.consts()->target(*labels.at(i), (address) constant_addr);
twisti@1915 3222 cb.consts()->relocate((address) constant_addr, relocInfo::internal_word_type);
twisti@1915 3223 }
twisti@1915 3224 }